DETAILED ACTION
This Office Action is in response to communications filed on April 20th, 2026 for Application No. 18/173,873, in which claims 1, 3, 8-15, and 17-20 are presented for examination. The amendments filed on April 20th, 2026 have been entered, where claims 1, 3, 8-15, and 17-18 are amended, claims 2, 4-7, and 16 are canceled, and claims 19-20 are newly added.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
Claims 1, 3, 8-15, and 17-20 are rejected under 35 U.S.C. 112(b) as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor regards as the invention.
Regarding Claim 1, the claim recites “such that pieces of input data that are greatly different from one another are selected as the candidate data” (ln. 60-61). The term “greatly” is a relative term which renders the claim indefinite. The term “greatly” is not defined by the claim, the specification does not provide a standard for ascertaining the requisite degree, and one of ordinary skill in the art would not be reasonably apprised of the scope of the invention. As a result, it is not clear how “different from one another” the “pieces of input data” that “are selected as the candidate data” should be. Therefore, the scope of the claim is indefinite. As a result, the claim is rejected. The claim should be modified to provide a standard for ascertaining the requisite degree of “greatly”.
Regarding claims 3 and 8, the claims are rejected because they are dependent on a claim that is rejected for indefiniteness.
Regarding claims 9-11, each of the claims recite “the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data output from the first position inside the . . . machine learning model” (Claim 9, ln. 21-25; Claim 10, ln. 15-18; and Claim 11, ln. 16-19). For each claim, it is unclear whether “which” is referencing “the second evaluation data” or “the first evaluation data”. As a result, the scope of each claim is indefinite because it is not clear whether the additional limitations regarding the “machine learning model” apply to the second evaluation data” or “the first evaluation data”. Therefore, the claims are rejected. The claim should be modified to clarify which “data” is being referenced by “which”.
Additionally, the claims are rejected because they are dependent on a claim that is rejected for indefiniteness.
Regarding claims 12-15, the claims are rejected because they are dependent on a claim that is rejected for indefiniteness.
Regarding claims 17-18, the claims recite “such that pieces of input data that are greatly different from one another are selected as the candidate data” (Claim 17, ln. 56-57; Claim 18, ln. 55-56), which are indefinite for substantially the same reasoning as articulated in regard to the rejection of Claim 1 above. As a result, the claims are similarly rejected and should be amended in a similar manner.
Regarding claims 19-20, the claims are rejected because they are dependent on a claim that is rejected for indefiniteness.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 3, 8-15, and 17-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract ideas without significantly more.
Regarding Claim 1:
Step 1: Claim 1 is a machine claim. Therefore, claims 1, 3, 8-15, and 19-20 are directed to a statutory category of eligible subject matter.
Step 2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the "Mental Processes" grouping of abstract ideas. Here, elements of the claimed subject matter are mental processes. Specifically, the claim recites
“select a plurality of pieces of learning data . . . from among a plurality of pieces of input data” (mental process – amounts to exercising judgment to form an opinion on known or observed information, which may be aided by pen and paper);
“generating pieces of input data arranged on a time-series basis based on observation results obtained by observing surroundings” (mental process – amounts to exercising judgement to form an opinion based on observation, which may be aided by pen and paper);
“inference processing on the respective pieces of input data on a time- series basis . . . obtained by performing the inference processing on the time-series basis” (mental process – amounts to exercising judgment to evaluate observed data, which may be aided by pen and paper);
“calculating a first evaluation value representing effectiveness of each of the pieces of input data . . . based on a first evaluation standard” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to other known or observed data and/or standards with specific constraints, which may be aided by pen and paper);
“selecting whether each of the pieces of input data is included in a plurality of pieces of candidate data by comparing the first evaluation value of each of the pieces of input data with a first standard value” (mental process – amounts to exercising judgment to form an opinion on known or observed information, with reference to other known or observed data and/or standards with specific constraints, which may be aided by pen and paper);
“calculating a second evaluation value indicating effectiveness of each of the pieces of candidate data . . . based on a second evaluation standard, the second evaluation standard being different from the first evaluation standard” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to other known or observed data and/or standards with specific constraints, which may be aided by pen and paper);
“selecting whether each of the pieces of candidate data is included in the pieces of learning data by comparing the second evaluation value of each of the pieces of candidate data with a second standard value” (mental process – amounts to exercising judgment to form an opinion on known or observed information, with reference to other known or observed data and/or standards with specific constraints, which may be aided by pen and paper);
“in the calculating of the first evaluation value, as the first evaluation value for first input data among the pieces of input data, calculating a value representing a degree of a difference between the first input data and k pieces of the candidate data . . . k being an integral number equal to or larger than 1, such that pieces of input data that are greatly different from one another are selected as the candidate data” (mental process – amounts to evaluating observed data to form an opinion on selection, which may be aided by pen and paper);
“in the selecting of whether each of the pieces of input data is included in the pieces of candidate data, when the first evaluation value for the first input data is larger than the first standard value, selecting that the first input data is included in the pieces of candidate data” (mental process - amounts to evaluating observed data to form an opinion on selection, with reference to a known standard, which may be aided by pen and paper); and
“in the calculating of the second evaluation value, calculating the second evaluation value for first candidate data among the pieces of candidate data by analyzing an inference result or an intermediate result” (mental process – amounts to evaluating observed data to form an opinion on selection, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“A machine learning system configured to . . . for causing a first machine learning model to perform learning . . . wherein the first hardware processor of the first information processing device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea);
“the machine learning system comprising: the machine learning system comprising: a first information processing device comprising at least one first hardware processor; and a second information processing device comprising at least one second hardware processor and connected to the first information processing device via a network. . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the first information processing device” (outputting, transmitting, and receiving data amounts to extra-solution activity because transmission of data over a network is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“A machine learning system configured to . . . for causing a first machine learning model to perform learning . . . wherein the first hardware processor of the first information processing device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept);
“the machine learning system comprising: the machine learning system comprising: a first information processing device comprising at least one first hardware processor; and a second information processing device comprising at least one second hardware processor and connected to the first information processing device via a network. . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the first information processing device” (transmitting data over a network is well‐understood, routine, and conventional, see Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362; see also buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014); therefore the limitation, which is recited with a high level of generality, remains insignificant extra-solution activity even upon reconsideration).
For the reasons above, Claim 1 is rejected as being directed to an abstract idea without significantly more. This rejection applies equally to dependent claims 3, 8-15, and 19-20. The additional limitations of the dependent claims are addressed below.
Regarding Claim 3:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 3 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the calculating of the first evaluation value . . . calculates, as the first evaluation value, a value according to a time difference between . . . time of the first input data and . . . time of second input data that is selected as one of the pieces of candidate data . . . among one or more pieces of the second input data” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to known or observed times associated with the information, which may be aided by pen and paper) and
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein . . . the first hardware processor” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea);
“immediately before the first input data . . . different from the first input data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea); and
“acquisition . . . acquisition” (in the event that the recitations of acquisition times implicitly required data gathering, gathering of data is extra-solution activity because it is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein . . . the first hardware processor” (mere instructions to apply the exception using generic computer components does not provide an inventive concept);
“immediately before the first input data . . . different from the first input data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept); and
“acquisition . . . acquisition” (data gathering is well-understood, routine and conventional, see OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93; therefore the limitation, which, at most, is implicitly recited with a high level of generality, remains insignificant extra-solution activity even upon reconsideration).
Accordingly, Claim 3 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 8:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 8 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“classifies input data into any of a plurality of classes . . . and in calculating of the second evaluation value . . . calculates, as the second evaluation value, a value according to a degree of difference representing a difference between a classification probability of a class into which the first candidate data is classified as belonging among the classes and one or a plurality of classification probabilities for one or a plurality of classes into which the first candidate data is classified as not belonging among the classes” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to a difference between known or determined data and descriptions associated with the information, which may be aided by pen and paper) and
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein the first machine learning model . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“acquires a classification probability of belonging to each of the classes” (acquiring classification probabilities is mere extra-solution activity because it is data gathering that is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein the first machine learning model . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“acquires a classification probability of belonging to each of the classes” (data gathering is well-understood, routine and conventional, see OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93; therefore the limitation, which is recited with a high level of generality, so remains insignificant extra-solution activity even upon reconsideration).
Accordingly, Claim 8 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 9:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 9 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the calculating of the second evaluation value . . . calculates, as the second evaluation value, a value according to a degree of difference representing a difference between first evaluation data . . . and second evaluation data” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to known or observed differences associated with the information, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . the first hardware processor . . . according to the first machine learning model . . . the second hardware processor . . . according to the first machine learning model . . . output from a first position inside the first machine learning model obtained by inputting the first candidate data to the first machine learning model . . . output from the first position inside the first machine learning model obtained by inputting the first candidate data to the first machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“performs arithmetic processing with first arithmetic accuracy, . . . performs arithmetic processing with second arithmetic accuracy higher than the first arithmetic accuracy . . . the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . the first hardware processor . . . according to the first machine learning model . . . the second hardware processor . . . according to the first machine learning model . . . output from a first position inside the first machine learning model obtained by inputting the first candidate data to the first machine learning model . . . output from the first position inside the first machine learning model obtained by inputting the first candidate data to the first machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“performs arithmetic processing with first arithmetic accuracy, . . . performs arithmetic processing with second arithmetic accuracy higher than the first arithmetic accuracy . . . the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 9 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 10:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 10 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in calculating of the second evaluation value . . . calculates, as the second evaluation value, a value according to a degree of difference representing a difference between first evaluation data . . . and second evaluation data . . . that is obtained by partially changing the first candidate data” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to an altered version of the information, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting data . . . to the first machine learning model . . . output from a first position inside the first machine learning model . . . output from the first position inside the first machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting data . . . to the first machine learning model . . . output from a first position inside the first machine learning model . . . output from the first position inside the first machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 10 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 11:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 11 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the calculating of the second evaluation value . . . calculates, as the second evaluation value, a value according to a degree of difference representing a difference between first evaluation data . . . and second evaluation” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to other known or observed information, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting the first candidate data to a second machine learning model that is obtained by partially changing the first machine learning model,. . . output from a first position inside the first machine learning model . . . output from the first position inside the first machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting the first candidate data to a second machine learning model that is obtained by partially changing the first machine learning model,. . . output from a first position inside the first machine learning model . . . output from the first position inside the first machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“the first evaluation data includes at least one of output data of the first machine learning model and intermediate data . . . and the second evaluation data includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 11 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 12:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 12 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the calculating of the second evaluation value . . . calculates, as the second evaluation value, a value representing variation among a plurality of pieces of output data” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to known or determined variance among the observed information, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to a plurality of machine learning models” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“and the pieces of output data are a plurality of inference results . . . learned with learning parameters different from learning parameters of the first machine learning model” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein . . . the second hardware processor . . . obtained by inputting the first candidate data to a plurality of machine learning models” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“and the pieces of output data are a plurality of inference results . . . learned with learning parameters different from learning parameters of the first machine learning model” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 12 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 13:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 13 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the calculating of the second evaluation value . . . calculates, as the second evaluation value, a value based on a degree of difference representing a difference between first output data and each of one or more pieces of second output data” (mental process – amounts to exercising judgment to determine a value associated with known or observed information, with reference to known or determined differences between the observed information, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting the first candidate data to one or more machine learning models” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea) and
“the first output data is an inference result . . . and the one or more pieces of second output data are respectively one or more inference results . . . learned with learning parameters different from learning parameters of the first machine learning model” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein the second hardware processor . . . obtained by inputting the first candidate data to the first machine learning model . . . obtained by inputting the first candidate data to one or more machine learning models” (mere instructions to apply the exception using generic computer components does not provide an inventive concept) and
“the first output data is an inference result . . . and the one or more pieces of second output data are respectively one or more inference results . . . learned with learning parameters different from learning parameters of the first machine learning model” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 13 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 14:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 14 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“makes a probability of selecting, as the candidate data, the input data acquired in a first time range . . . that is determined with the first time range as a standard, to be higher than a probability of selecting, as the candidate data” (mental process – amounts to exercising judgement to form opinion that information within a previously determined time range should be prioritized over information in another time range).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein the second hardware processor further . . . indicating that corresponding input data is selected as the learning data . . . each time the learning data is selected . . . that is after a first time being a time of the input data indicated by the employment information . . . the input data acquired in a time range other than the first time range” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea) and
“transmits employment information . . . to the first hardware processor of the information processing device . . . and the first hardware processor receives the employment information” (transmitting and receiving data amounts to extra-solution activity because transmission of data over a network is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein the second hardware processor further . . . indicating that corresponding input data is selected as the learning data . . . each time the learning data is selected . . . that is after a first time being a time of the input data indicated by the employment information . . . the input data acquired in a time range other than the first time range” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept) and
“transmits employment information . . . to the first hardware processor of the information processing device . . . and the first hardware processor receives the employment information” (transmitting data over a network is well‐understood, routine, and conventional, see Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362; see also buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014); therefore the limitation, which is recited with a high level of generality, remains insignificant extra-solution activity even upon reconsideration).
Accordingly, Claim 14 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 15:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 15 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“in the selecting of whether each of the pieces of input data is included in the pieces of candidate data . . . determines whether to select each of the pieces of input data as a candidate based on a corresponding first evaluation value on a time-series basis” (mental process – amounts to exercising judgment to form an opinion on known or observed information, with reference to other known or observed data on a sequential basis, which may be aided by pen and paper); and
“in the selecting of whether each of the pieces of candidate data is included in the pieces of learning data . . . determines whether each of the pieces of candidate data is included in the pieces of learning data based on a corresponding second evaluation value on a time-series basis” (mental process – amounts to exercising judgment to form an opinion on known or observed information, with reference to other known or observed data on a sequential basis, which may be aided by pen and paper).
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“the first hardware processor . . . the second hardware processor” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“the first hardware processor . . . the second hardware processor” (mere instructions to apply the exception using generic computer components does not provide an inventive concept).
Accordingly, Claim 15 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 17:
Step 1: Claim 17 is a machine claim. Therefore, it is directed to a statutory category of eligible subject matter.
Step 2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the "Mental Processes" grouping of abstract ideas. Here, the claim recites elements that are substantially the same as the limitations of Claim 1. As a result, and as elaborated above, these limitations are abstract ideas because they are mental processes.
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“An edge device in a machine learning system . . . for causing a first machine learning model to perform learning . . . wherein . . . the first hardware processor of the edge device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea);
“that comprises the edge device and an information processing device connected to the edge device via a network . . . the edge device comprises at least one first hardware processor, the information processing device comprises at least one second hardware processor . . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the edge device” (outputting, transmitting, and receiving data amounts to extra-solution activity because transmission of data over a network is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“An edge device in a machine learning system . . . for causing a first machine learning model to perform learning . . . wherein . . . the first hardware processor of the edge device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept);
“that comprises the edge device and an information processing device connected to the edge device via a network . . . the edge device comprises at least one first hardware processor, the information processing device comprises at least one second hardware processor . . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the edge device” (transmitting data over a network is well‐understood, routine, and conventional, see Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362; see also buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014); therefore the limitation, which is recited with a high level of generality, remains insignificant extra-solution activity even upon reconsideration).
For the reasons above, Claim 17 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 18:
Step 1: Claim 18 is a machine claim. Therefore, it is directed to a statutory category of eligible subject matter.
Step 2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the "Mental Processes" grouping of abstract ideas. Here, the claim recites elements that are substantially the same as the limitations of Claim 1. As a result, and as elaborated above, these limitations are abstract ideas because they are mental processes.
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“An information processing device in a machine learning system . . . for causing a first machine learning model to perform learning . . . wherein . . . the first hardware processor of the edge device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea);
“that comprises the edge device and an information processing device connected to the edge device via a network . . . the edge device comprises at least one first hardware processor, the information processing device comprises at least one second hardware processor . . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the edge device” (outputting, transmitting, and receiving data amounts to extra-solution activity because transmission of data over a network is incidental to the claimed subject matter).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“An information processing device in a machine learning system . . . for causing a first machine learning model to perform learning . . . wherein . . . the first hardware processor of the edge device performs . . . based on the first machine learning model . . . the second hardware processor of the second information processing device performs . . . training the first machine learning model . . . updating the parameters used for the inference processing performed by the first hardware processor . . . wherein the first hardware processor performs . . . the second hardware processor performs . . . obtained by inputting the first candidate data to a machine learning model” (mere instructions to apply the exception using generic computer components does not provide an inventive concept);
“that comprises the edge device and an information processing device connected to the edge device via a network . . . the edge device comprises at least one first hardware processor, the information processing device comprises at least one second hardware processor . . . when being used for learning of the first machine learning model . . . when being used for learning of the first machine learning model . . . by using the pieces of learning data . . . that are set in the first machine learning model trained . . . immediately before the first input data among the pieces of candidate data” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept); and
“outputting an inference result . . . transmitting each of the pieces of candidate data to the second information processing device via the network . . . receiving each of the pieces of candidate data from the first information processing device via the network . . . transmitting parameters . . . to the first hardware processor of the edge device” (transmitting data over a network is well‐understood, routine, and conventional, see Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362; see also buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014); therefore the limitation, which is recited with a high level of generality, remains insignificant extra-solution activity even upon reconsideration).
For the reasons above, Claim 18 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 19:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 19 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“generates the first evaluation value that becomes a larger value as the difference between the first input data and the k pieces of the candidate data is larger” (mental process – amounts to exercising judgement to form opinions on known or observed information, such that the evaluation criteria conforms with a specific trend, which may be aided by pen and paper);
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein the first hardware processor” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein the first hardware processor” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept).
Accordingly, Claim 19 is rejected as being directed to an abstract idea without significantly more.
Regarding Claim 20:
Step 2A Prong 1: See the rejection of Claim 19 above, which Claim 20 depends on. Here, the claim recites additional limitations that are mental processes. Specifically, the claim recites
“calculates, as the first evaluation value, a Sum of Absolute Difference (SAD) or a Sum of Squared Difference (SSD)” (mental process – amounts to exercising judgement to form opinions on known or observed information, which may be aided by pen and paper);
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
The claim recites the additional elements:
“wherein the first hardware processor” (amounts to merely generally linking the use of the judicial exception to a particular technological environment or field of use, which do not impose any meaningful limits on practicing the abstract idea) and
“when each of the plurality of pieces of input data is image data” (amounts to mere instructions to apply the judicial exception on generic and unspecialized computer components, which do not impose any meaningful limits on practicing the abstract idea).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
The claim recites the additional elements:
“wherein the first hardware processor” (merely generally linking the use of the judicial exception to a particular technological environment or field of use does not provide an inventive concept) and
“when each of the plurality of pieces of input data is image data” (mere instructions to apply the exception using generic computer components does not provide an inventive concept).
Accordingly, Claim 20 is rejected as being directed to an abstract idea without significantly more.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 15, and 17-19 are rejected under 35 U.S.C. 103 as being unpatentable over Karpathy et al. (hereinafter Karpathy) (Patent Pub. No. US 2021/0271259 A1) in view of Kanno et al. (hereinafter Kanno) (Patent Pub. No. US 2021/0004723 A1).
Regarding Claim 1, Karpathy teaches a machine learning system configured to select a plurality of pieces of learning data for causing a first machine learning model to perform learning from among a plurality of pieces of input data (Abstract, “Systems and methods for obtaining training data . . . includes . . . applying a neural network to the sensor data. A trigger classifier is applied to an intermediate result of the neural network to determine a classifier score for the sensor data. Based at least in part on the classifier score, a determination is made whether to transmit via a computer network at least a portion of the sensor data. Upon a positive determination, the sensor data is transmitted and used to generate training data”, where “Systems” that include “applying a neural network” are machine learning systems, which are configured to select, “determination is made whether to transmit . . . [and use] to generate training data”, a plurality of pieces of learning data of learning data from a plurality of pieces of input data, where the selected “portion of the sensor data” are the learning data, selected from the input data of all “sensor data”; see also Para. [0081], “sensor data received is processed to create training data for training a machine learning model”, where the “training data” is used for causing a first “machine learning model” to perform learning through “training”),
the machine learning system comprising: a first information processing device comprising at least one first hardware processor; and a second information processing device comprising at least one second hardware processor and connected to the first information processing device via a network, wherein the first hardware processor of the first information processing device performs
(Para. [0007], “FIG. 1B is a block diagram illustrating one embodiment of a system for generating training data” and FIG. 1B, where the machine learning system, as indicated by the inclusion of a “Deep Learning System 700” component, includes a vehicle “102” as a first device and a “Training Data Generating System 120” as a second device, which are connected via the “Network”; see also Para.[0023], “classifiers may be uploaded to a computer system within a vehicle, such that the classifier may be used to recognize specific image features or objects associated with the classifiers. The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system and used as training data for neural network systems”, where both the “vehicle”, which as discussed above is “102”, and the “server”, which is 120, see Para. [0080], “a computer server (e.g., the training data generation system 120)”, are information processing devices because the “vehicle” processes “image features or objects”, which requires at least one first processor, and the “server” processes “training data”, which requires at least one second processor; see also Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Fig. 4, where, as discussed above, the “vehicle” is the first information processing device, which executes the functionality of, “Determine Trigger Classification Score 411”, “Score Exceeds Threshold And Conditions Met 413”, and “Transmit Identified Sensor Data 415”, which requires at least one first hardware processor, see generally Para. [0103], “The various aspects, embodiments, implementations or features of the described embodiments can be used separately or in any combination. Various aspects of the described embodiments can be implemented by software, hardware or a combination of hardware and software . . . The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable code is stored and executed in a distributed fashion”, which could be within a computer system or “distributed over network-coupled computer systems”; see also Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, which as discussed above is the second information processing device, performs the functionality of “FIG. 5[’s] . . . flow diagram”; see also Fig. 5, where the functionality of “Receive Sensor Data Meeting Trigger Conditions 501”, “Convert Sensor Data Into Training Data 503”, and “Prepare Training And Validation Data Sets 505” and “Train Machine Learning Model 507”, each require at least a second hardware processor, see generally Para. [0103], “The various aspects, embodiments, implementations or features of the described embodiments can be used separately or in any combination. Various aspects of the described embodiments can be implemented by software, hardware or a combination of hardware and software . . . The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable code is stored and executed in a distributed fashion”, where the units of “hardware and software” could be within a computer system or “distributed over network-coupled computer systems”):
generating pieces of input data arranged on a time-series basis based on observation results obtained by observing surroundings (Fig. 2, where, the first hardware processor required to execute the “Receive Sensor Data 201” and “Perform Data Pre-processing 203” functionality to “capture sensor data”, by observing surroundings results, “the surrounding environment”, to generate pieces of input data, “the captured image is provided for deep learning analysis”, see Para. [0054], “At 201, sensor data is received. For example, a vehicle equipped with sensors captures sensor data and provides the sensor data to a neural network running on the vehicle . . . to capture data of the surrounding environment . . . the captured image is provided for deep learning analysis”; see also Para. [0021], “the sensor information may be captured in the normal course of operation of the vehicles. The sensor information may be used by the vehicles for certain automated driving features, such as lane navigation” and Para. [0080], “the sensor data transmitted includes metadata. Examples of metadata may include the time of data, a timestamp . . . compression of multiple images of sensor data is performed and a series of sensor data is transmitted together”, where the generated input data, “sensor data transmitted”, includes “timestamp[s]”, can be organized as “a series of sensor data”, and is “captured” over the course of time, “normal course of operation”, which is within the broadest reasonable interpretation of arranged on a time-series basis);
inference processing on the respective pieces of input data on a time- series basis based on the first machine learning model; outputting an inference result obtained by performing the inference processing on the time-series basis (Fig. 2, where, as discussed above, the first hardware processor is required to execute the “Initiate Deep Learning Analysis 205”, “Perform Data Post-processing 211”, and “Provide Results To Vehicle Control 213” functionality, which is configured to perform inference processing on the respective pieces of input data, “deep learning analysis of the sensor data”, based on the first machine learning model, “a convolutional neural network (CNN)”, which must be on a time-series basis to “identify . . . moving vehicles”, see Para. [0056], “At 205, deep learning analysis of the sensor data is initiated . . . using a neural network such as a convolutional neural network (CNN). In various embodiments, the machine learning model is trained offline and installed onto the vehicle for performing inference on the sensor data. For example, the model may be trained to identify road lane lines, obstacles, pedestrians, moving vehicles, parked vehicles, drivable space, etc., as appropriate”, and output inference “results”, which must be in a time series basis to allow for “autonomous driving”, see Para. [0060], “At 213, the results of the deep learning analysis are provided to vehicle control. For example, the results are used by a vehicle control module to control the vehicle for autonomous driving”; see also Para. [0030], “A new machine learning model is trained using the newly curated data set to improve the autonomous vehicle neural network, and is then deployed to vehicles as an update to the autonomous vehicle system”, where the first machine learning model, “the autonomous vehicle neural network”, is the model “deployed to [the] vehicles”);
calculating a first evaluation value representing effectiveness of each of the pieces of input data when being used for learning of the first machine learning model based on a first evaluation standard (Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Determine Trigger Classification Score 411” functionality, which is configured to calculate a first evaluation value, “classification score”, see Para. [0077], “At 411, a trigger classifier score is determined”; see also Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the “trigger classifier” determines the trigger score as a representation of the effectiveness of each piece of input data in machine learning of the first model, “analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases”, based on a first evaluation standard, whether “the sensor data” “warrant retraining”; see also Para. [0102], “trigger classifier module 713 determines a classifier score for a data captured by one or more sensors of sensors”, where each “classifier score” is for each of the pieces of data because it is for “a data captured”);
selecting whether each of the pieces of input data is included in a plurality of pieces of candidate data by comparing the first evaluation value of each of the pieces of input data with a first standard value (Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Score Exceeds Threshold And Conditions Met 413” functionality, which is configured to compare the first evaluation value of each of the pieces of input data, which as discussed above is the “classifier score”, with a first standard value, “threshold value”, which determines whether the pieces of input data associated with the “classifier score” will be selected for inclusion in “continue[d]” “processing” at “415”, see Para. [0078], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415”; see also Para. [0060], “the intermediate results of the deep learning analysis at 205 are utilized for identifying training data at 207 and transmitting the identified sensor data at 209”, where the plurality of pieces of “sensor data” “identified” for inclusion in future processing are candidate data for “training data”); and
transmitting each of the pieces of candidate data to the second information processing device via the network (Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Transmit Identified Sensor Data 415” functionality, which is configured to “transmit” each of the pieces of candidate data, “identified sensor data”, to the second information processing device, “server (e.g., the training data generation system 120)”, see Para. [0080], “At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”; see also Para. [0044], “the vehicle may transmit the sensor data 108 over a network”, where the “transmi[ssion]” is “over a network”), and
the second hardware processor of the second information processing device performs (Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, which as discussed above is the second information processing device, performs the functionality of “FIG. 5[’s] . . . flow diagram”; see also Fig. 5, where the functionality of “Receive Sensor Data Meeting Trigger Conditions 501”, “Convert Sensor Data Into Training Data 503”, and “Prepare Training And Validation Data Sets 505” and “Train Machine Learning Model 507”, each require at least a second hardware processor, see generally Para. [0103], “The various aspects, embodiments, implementations or features of the described embodiments can be used separately or in any combination. Various aspects of the described embodiments can be implemented by software, hardware or a combination of hardware and software . . . The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable code is stored and executed in a distributed fashion”, where the units of “hardware and software” could be within a computer system or “distributed over network-coupled computer systems”):
receiving each of the pieces of candidate data from the first information processing device via the network (Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Receive Sensor Data Meeting Trigger Conditions 501” functionality, which is configured to “receive” pieces of “sensor data”, see Para. [0082], “At 501, sensor data meeting trigger conditions is received”; see also Para. [0081], “the sensor data is received using the process of FIG. 4”, where, as discussed above, “the process of FIG. 4”, includes the transmission of each of the pieces of candidate data from the first information processing device; see also Para. [0093], “The sensor data that triggers retention for transmittal by trigger classifier module 713 is sent via network interface 711 . . . [by] the vehicle”, where the transmission is “via network”);
calculating a second evaluation value indicating effectiveness of each of the pieces of candidate data when being used for learning of the first machine learning model based on a second evaluation standard, the second evaluation standard being different from the first evaluation standard (Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which is output based on the configuration of the “machine learning model”, the second evaluation standard in this instance; see also Para. [0045], “The classifiers 110A-110N may, as an example, use classifier scores which cause transmission of a multitude of sensor data 108 to the outside system 120. For example, a portion of images transmitted to the system 120 may not include tires. In some embodiments, the entity may thus rapidly review and discard certain of the images”, where the second evaluation standard must be different from the first evaluation standard for “a portion of images transmitted to the system” to be “discard[ed]” based on the second evaluation standard, despite the determination of “sensor data” for “transmission” being based on the first evaluation standard, see Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”);
selecting whether each of the pieces of candidate data is included in the pieces of learning data by comparing the second evaluation value of each of the pieces of candidate data with a second standard value (Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Prepare Training And Validation Data Sets 505” and “Train Machine Learning Model 507” functionality, which is configured to select whether each of the pieces of candidate data, “the training data of 503”, is included in the pieces of learning data, “merged into existing training data sets”, by comparing “the training data of 503” with a second standard value, “a particular use case”, see Para. [0084] – [0085], “At 505 . . . the training data of 503 is merged into existing training data sets. For example, an existing training data set applicable for most use cases is merged with the newly converted training data for improved coverage of a particular use case. The newly converted training data is useful for improving the accuracy of the model in identifying the particular use case. At 507, a machine learning model is trained . . . using the data prepared at 505”; see also Para. [0011], “FIG. 5 is a flow diagram illustrating an embodiment of a process for deploying training data from data corresponding to use cases identified by a trigger classifier”, where the “use cases” must be determined in order to be “identified by a trigger classifier”, which occurs in advance of the “flow diagram” functionality of “FIG.5”; see also Para. [0083], “For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where using the output from “the machine learning model”, which as discussed above, is the second evaluation value, to “confirm whether the sensor data represents the targeted use case”, which as discussed above is the value, is withing the broadest reasonable interpretation of comparing the second evaluation value with the value);
training the first machine learning model by using the pieces of learning data (Para. [0030], “after uploading the sensor data is reviewed and annotated to create a new training data set that is used to improve the autonomous driving features of the vehicle. For example, the data may be annotated as positive samples of tunnel exits and may be used to supplement an original training data set that includes many more use cases. A new machine learning model is trained using the newly curated data set to improve the autonomous vehicle neural network”, where the second hardware processor is required to perform the “train[ing]” functionality, which uses the pieces of learning data, “a new training data set that is used” to train the first machine learning model, “A new machine learning model is trained using the newly curated data set”);
transmitting parameters that are set in the first machine learning model trained, to the first hardware processor of the first information processing device; and updating the parameters used for the inference processing performed by the first hardware processor (Para. [0030] and [0031], “A new machine learning model is trained using the newly curated data set to improve the autonomous vehicle neural network, and is then deployed to vehicles as an update to the autonomous vehicle system . . . The process can be performed remotely and dynamically, for example, using an over-the-air update, without requiring the vehicle be brought to a service location”, where the second hardware processor is required to perform the “deploy[ment] to vehicles”, which transmits, “remotely and dynamically, for example, using an over-the-air update”, parameters that are set in the first machine learning model trained, “A new machine learning model is trained”, to the first hardware processor of the first information processing device, “then deployed to vehicles as an update to the autonomous vehicle system”, updating the parameters, see Para. [0017], “During training, these examples may be used to adjust parameters of the neural network (e.g., weights, biases, and so on). Additionally, these examples may be used to adjust hyperparameters of the neural network (e.g., a number of layers)”, used for the inference processing performed by the first hardware processor, “to improve the autonomous vehicle neural network”; see also Fig. 2, where, as discussed above, the first hardware processor is required to execute the “Initiate Deep Learning Analysis 205”, “Perform Data Post-processing 211”, and “Provide Results To Vehicle Control 213” functionality, which is configured to perform inference processing on the respective pieces of input data, “deep learning analysis of the sensor data”, based on the first machine learning model, “a convolutional neural network (CNN)”, which must be on a time-series basis to “identify . . . moving vehicles”, see Para. [0056], “At 205, deep learning analysis of the sensor data is initiated . . . using a neural network such as a convolutional neural network (CNN). In various embodiments, the machine learning model is trained offline and installed onto the vehicle for performing inference on the sensor data. For example, the model may be trained to identify road lane lines, obstacles, pedestrians, moving vehicles, parked vehicles, drivable space, etc., as appropriate”),
wherein the first hardware processor performs: in the calculating of the first evaluation value, as the first evaluation value for first input data among the pieces of input data, calculating a value . . . such that pieces of input data that are greatly different from one another are selected as the candidate data (Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Determine Trigger Classification Score 411” functionality, which is configured to calculate a first evaluation value, “classification score”, see Para. [0077], “At 411, a trigger classifier score is determined”; see also Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the “trigger classifier” determines the trigger score as a representation of the effectiveness of each piece of input data in machine learning of the first model, “analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases”, based on a first evaluation standard, whether “the sensor data” “warrant retraining”; see also Para. [0102], “trigger classifier module 713 determines a classifier score for a data captured by one or more sensors of sensors”, where each “classifier score” is for each of the pieces of data because it is for “a data captured”, such that pieces of input data that are greatly different from one another are selected as the candidate data, see Para. [0023], “The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system . . . there may be a large number of vehicles being driven in disparate environments, which increases the likelihood of obtaining examples of hard to find ‘edge cases’ of certain features” and Para. [0037], “The training data may be enhanced by the inclusion of images of different tires on different roads. Additionally, the training data may be enhanced by images of different tires on different roads in different driving environments”); and
in the selecting of whether each of the pieces of input data is included in the pieces of candidate data, when the first evaluation value for the first input data is larger than the first standard value, selecting that the first input data is included in the pieces of candidate data (Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Score Exceeds Threshold And Conditions Met 413” functionality, which is configured to compare the first evaluation value of each of the pieces of input data, which as discussed above is the “classifier score”, with the first standard value, “threshold value”, which, if larger than the first standard value, determines that the pieces of input data associated with the “classifier score” will be selected for inclusion in “continue[d]” “processing” at “415”, see Para. [0078], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415”; see also Para. [0060], “the intermediate results of the deep learning analysis at 205 are utilized for identifying training data at 207 and transmitting the identified sensor data at 209”, where the plurality of pieces of “sensor data” “identified” for inclusion in future processing are candidate data for “training data”), and
the second hardware processor performs: in the calculating of the second evaluation value, calculating the second evaluation value for first candidate data among the pieces of candidate data . . . by inputting the first candidate data to a machine learning model (Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which, as also discussed above, is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires a calculated evaluation value, the second evaluation value in this instance, which is output based on the configuration of the “machine learning model”, the second evaluation standard in this instance, which receives the first candidate data as input, “data identified as potentially useful training data”).
Karpathy does not explicitly disclose . . . representing a degree of a difference between the first input data and k pieces of the candidate data immediately before the first input data among the pieces of candidate data, k being an integral number equal to or larger than 1 . . . by analyzing an inference result or an intermediate result obtained . . . .
However, Kanno teaches . . . [calculating an evaluation value] representing a degree of a difference between the first input data and k pieces of the candidate data (Para. [0036], “ the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference. Here, it is assumed that the selecting unit 4 sorts the training data in ascending order based on a value indicating the difference”, where the calculated evaluation value, the rank used to “sort . . . in ascending order”, is according to a degree of difference, “indicating the difference”, which represents a difference, the variability in “difference between a category determined for training data and correct answer data” between the candidate data, the “each of training data”, where “sort[ing] . . . in ascending order” requires a plurality of data, represented as an integral greater than 1, where one input is the first input data and the remaining are the k pieces)
immediately before the first input data among the pieces of candidate data, k being an integral number equal to or larger than 1 . . . (Para. [0036], “ the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference. Here, it is assumed that the selecting unit 4 sorts the training data in ascending order based on a value indicating the difference”, where, as discussed above, “sort[ing] . . . in ascending order” “each of training data”, requires a plurality of data, represented as an integral greater than 1, where one input is the first input data and the remaining are the k pieces, and where, in this instance, the first input data is the data “in [the] ascending order”, which is immediately after the k pieces, which are immediately before it)
. . . [and calculating an evaluation value for data] by analyzing an inference result or an intermediate result obtained [by inputting the data into a machine learning model] . . . (Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”, where the “value indicating the difference” between the inference “category” of the data and the “correct answer” for the data, is the evaluation value, which is calculated by analyzing the “category” inference of the “model”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the first hardware processor that calculates the first evaluation value, as the first evaluation value for first input data among the pieces of input data, such that pieces of input data that are greatly different from one another are selected as the candidate data and the second hardware processor that calculates the second evaluation value for first candidate data among the pieces of candidate data by inputting the first candidate data to a machine learning model of Karpathy with the calculating an evaluation value representing a degree of a difference between the first input data and k pieces of the candidate data immediately before the first input data among the pieces of candidate data, k being an integral number equal to or larger than 1 and calculating an evaluation value for data by analyzing an inference result or an intermediate result obtained by inputting the data into a machine learning model of Kanno in order to operationalize the selection operations of the first hardware processor using an evaluation value that clearly organizes data pieces based on whether each should be considered for further evaluation (Karpathy, Para. [0023], “there may be a large number of vehicles being driven in disparate environments, which increases the likelihood of obtaining examples of hard to find ‘edge cases’ of certain features” and Karpathy, Para. [0037], “The training data may be enhanced by the inclusion of images of different tires on different roads. Additionally, the training data may be enhanced by images of different tires on different roads in different driving environments”; see generally Kanno, Para. [0036], “the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference”, where associating data pieces with integral values allows for clear metrics that can easily be compared against the data immediately before and after it, which could easily be utilized for any comparison-based selection, such as selection of pieces that are greatly different from one another to allow for “obtaining examples of hard to find ‘edge cases’ of certain features” like “images of different tires on different roads in different driving environments”), which allows for greater specificity of model training and, as a result, increased inference accuracy of trained models on specific use cases (see Karpathy, Para. [0023], “there may be a large number of vehicles being driven in disparate environments, which increases the likelihood of obtaining examples of hard to find ‘edge cases’ of certain features” and Karpathy, Para. [0037], “The training data may be enhanced by the inclusion of images of different tires on different roads. Additionally, the training data may be enhanced by images of different tires on different roads in different driving environments”; see also Kanno, Para. [0003], “When training data includes training data not having a characteristic affecting determination of a category, accuracy of determining a learned model is reduced, or learning of a model is adversely affected. Therefore, it is necessary to remove training data not having a characteristic affecting determination of a category from collected training data”) by allowing only transmission of candidate data pieces to the server which conform a desired range (Kanno, Para. [0038], “Therefore, in the training data sorted in ascending order based on a value indicating the difference, higher training data can be said to be appropriate training data, and lower training data can be said to be inappropriate training data”; Karpathy, Para. [0080], “At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”) and in order to convert model outputs into evaluation metrics that clearly determine whether candidate data should be used as training data (Kanno, Para. [0038], “It can be said that training data having a small difference from correct answer data is appropriate as training data used for learning the first model. In addition, it can be said that training data having a large difference from correct answer data is inappropriate as training data used for learning the first model”, where converting model inferences to a usable value allows for clear metrics that can easily be compared, see Kanno, Para. [0036], “the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference”), which allows for improved model training and increased inference accuracy of trained models (Kanno, Para. [0003], “When training data includes training data not having a characteristic affecting determination of a category, accuracy of determining a learned model is reduced, or learning of a model is adversely affected. Therefore, it is necessary to remove training data not having a characteristic affecting determination of a category from collected training data”).
Regarding Claim 15, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the selecting of whether each of the pieces of input data is included in the pieces of candidate data (Karpathy, Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Score Exceeds Threshold And Conditions Met 413” functionality, which is configured to compare the first evaluation value of each of the pieces of input data, which as discussed above is the “classifier score”, with a first standard value, “threshold value”, which determines whether the pieces of input data associated with the “classifier score” will be selected for inclusion in “continue[d]” “processing” at “415”, see Karpathy, Para. [0078], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415”; see also Karpathy, Para. [0060], “the intermediate results of the deep learning analysis at 205 are utilized for identifying training data at 207 and transmitting the identified sensor data at 209”, where the plurality of pieces of “sensor data” “identified” for inclusion in future processing are candidate data for “training data”),
the first hardware processor determines whether to select each of the pieces of input data as a candidate based on a corresponding first evaluation value on a time-series basis (Karpathy, Fig. 4, where the first hardware processor is required to execute the “Score Exceeds Threshold And Conditions Met 413” functionality, which, as also discussed above, is configured to compare the first evaluation value of each of the pieces of input data, which as discussed above is the “classifier score”, with a value, “threshold value”, which determines whether the pieces of input data associated with the “classifier score” will be selected for inclusion in “continue[d]” “processing” at “415”, on a time-series basis, “within the last 10 minutes may be retained”, see Karpathy, Para. [0078], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415 . . . only sensor data with the highest score from the same location within the last 10 minutes may be retained as potential data”; see generally Karpathy, Para. [0103], “The various aspects, embodiments, implementations or features of the described embodiments can be used separately or in any combination. Various aspects of the described embodiments can be implemented by software, hardware or a combination of hardware and software . . . The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable code is stored and executed in a distributed fashion”, where the units of “hardware and software” could be within a computer system or “distributed over network-coupled computer systems”), and
in the selecting of whether each of the pieces of candidate data is included in the pieces of learning data (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Prepare Training And Validation Data Sets 505” and “Train Machine Learning Model 507” functionality, which is configured to select whether each of the pieces of candidate data, “the training data of 503”, is included in the pieces of learning data, “merged into existing training data sets”, by comparing “the training data of 503” with a second standard value, “a particular use case”, see Karpathy, Para. [0084] – [0085], “At 505 . . . the training data of 503 is merged into existing training data sets. For example, an existing training data set applicable for most use cases is merged with the newly converted training data for improved coverage of a particular use case. The newly converted training data is useful for improving the accuracy of the model in identifying the particular use case. At 507, a machine learning model is trained . . . using the data prepared at 505”; see also Karpathy, Para. [0011], “FIG. 5 is a flow diagram illustrating an embodiment of a process for deploying training data from data corresponding to use cases identified by a trigger classifier”, where the “use cases” must be determined in order to be “identified by a trigger classifier”, which occurs in advance of the “flow diagram” functionality of “FIG.5”; see also Karpathy, Para. [0083], “For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where using the output from “the machine learning model”, which as discussed above, is the second evaluation value, to “confirm whether the sensor data represents the targeted use case”, which as discussed above is the value, is withing the broadest reasonable interpretation of comparing the second evaluation value with the value),
the second hardware processor determines whether each of the pieces of candidate data is included in the pieces of learning data based on a corresponding second evaluation value on a time-series basis (Karpathy, Fig. 5, where, as discussed above, the second software processor is required to execute the “Prepare Training And Validation Data Sets 505” and “Train Machine Learning Model 507” functionality, which, as also discussed above, is configured to select whether each of the pieces of candidate data, “the training data of 503”, is included in the pieces of learning data, “merged into existing training data sets”, by comparing “the training data of 503” with a value, “a particular use case”, see Karpathy, Para. [0084] – [0085], “At 505 . . . the training data of 503 is merged into existing training data sets. For example, an existing training data set applicable for most use cases is merged with the newly converted training data for improved coverage of a particular use case. The newly converted training data is useful for improving the accuracy of the model in identifying the particular use case. At 507, a machine learning model is trained . . . using the data prepared at 505”; see also Karpathy, Para. [0011], “FIG. 5 is a flow diagram illustrating an embodiment of a process for deploying training data from data corresponding to use cases identified by a trigger classifier”, where the “use cases” must be determined in order to be “identified by a trigger classifier”, which occurs in advance of the “flow diagram” functionality of “FIG.5”; see also Karpathy, Para. [0083], “For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where using the output from “the machine learning model”, which as discussed above, is the second evaluation value, to “confirm whether the sensor data represents the targeted use case”, which as discussed above is the value, is withing the broadest reasonable interpretation of comparing the second evaluation value with the value; see also Karpathy, Para. [0021], “the sensor information may be captured in the normal course of operation of the vehicles. The sensor information may be used by the vehicles for certain automated driving features, such as lane navigation”, Karpathy, Para. [0080], “the sensor data transmitted includes metadata. Examples of metadata may include the time of data, a timestamp . . . compression of multiple images of sensor data is performed and a series of sensor data is transmitted together”, and Karpathy, Para. [0078], “only sensor data with the highest score from the same location within the last 10 minutes may be retained as potential data”, where determinations of the processor are on a time-series basis are on a time-series basis because, as discussed above, it selects time-series data and receives the time-series data on a time-series basis).
Regarding Claim 17, Karpathy teaches an edge device in a machine learning system that comprises the edge device and an information processing device connected to the edge device via a network (Para. [0007], “FIG. 1B is a block diagram illustrating one embodiment of a system for generating training data” and FIG. 1B, where the machine learning system, as indicated by the inclusion of a “Deep Learning System 700” component, includes a vehicle “102” as a first device and a “Training Data Generating System 120” as a second device, which are connected via the “Network”; see also Para.[0023], “classifiers may be uploaded to a computer system within a vehicle, such that the classifier may be used to recognize specific image features or objects associated with the classifiers. The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system and used as training data for neural network systems”, where the “vehicle”, which as discussed above is “102”, is an edge device because it is at the edge of the “system”, distant from the “central server system”, and where the “server”, which is 120, see Para. [0080], “a computer server (e.g., the training data generation system 120)”, is an information processing device because it processes “training data”),
and selects a plurality of pieces of learning data for causing a first machine learning model to perform learning from among a plurality of pieces of input data (Abstract, “Systems and methods for obtaining training data . . . includes . . . applying a neural network to the sensor data. A trigger classifier is applied to an intermediate result of the neural network to determine a classifier score for the sensor data. Based at least in part on the classifier score, a determination is made whether to transmit via a computer network at least a portion of the sensor data. Upon a positive determination, the sensor data is transmitted and used to generate training data”, where the “Systems” selects, “determination is made whether to transmit . . . [and use] to generate training data”, a plurality of pieces of learning data of learning data from a plurality of pieces of input data, where the selected “portion of the sensor data” are the learning data, selected from the input data of all “sensor data”; see also Para. [0081], “sensor data received is processed to create training data for training a machine learning model”, where the “training data” is used for causing a first “machine learning model” to perform learning through “training”; see also Para. [0087], “the vehicle may determine classifier scores”, where the “vehicle” edge device, in combination other elements of the system, performs the selecting because, as a nonexclusive example, it “determine[s] classifier score”), wherein
the edge device comprises at least one first hardware processor, the information processing device comprises at least one second hardware processor, the first hardware processor of the edge device performs (Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Fig. 4, where, as discussed above, the “vehicle” is the edge device, which is the first information processing device, Para.[0023], “classifiers may be uploaded to a computer system within a vehicle, such that the classifier may be used to recognize specific image features or objects associated with the classifiers. The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system and used as training data for neural network systems”, where the “vehicle”, which as discussed above is “102”, is the first information processing device because the “vehicle” processes “image features or objects”, which requires at least one first processor to execute the functionality of, “Determine Trigger Classification Score 411”, “Score Exceeds Threshold And Conditions Met 413”, and “Transmit Identified Sensor Data 415” see generally Para. [0103], “The various aspects, embodiments, implementations or features of the described embodiments can be used separately or in any combination. Various aspects of the described embodiments can be implemented by software, hardware or a combination of hardware and software . . . The computer readable medium can also be distributed over network-coupled computer systems so that the computer readable code is stored and executed in a distributed fashion”, which could be within a computer system or “distributed over network-coupled computer systems”): . . .
receiving each of the pieces of candidate data from the edge device via the network (Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Receive Sensor Data Meeting Trigger Conditions 501” functionality, which is configured to “receive” pieces of “sensor data”, see Para. [0082], “At 501, sensor data meeting trigger conditions is received”; see also Para. [0081], “the sensor data is received using the process of FIG. 4”, where, as discussed above, “the process of FIG. 4”, includes the transmission of each of the pieces of candidate data from the edge device; see also Para. [0093], “The sensor data that triggers retention for transmittal by trigger classifier module 713 is sent via network interface 711 . . . [by] the vehicle”, where the transmission is “via network”) . . .
transmitting parameters that are set in the first machine learning model trained, to the first hardware processor of the edge device (Para. [0030] and [0031], “A new machine learning model is trained using the newly curated data set to improve the autonomous vehicle neural network, and is then deployed to vehicles as an update to the autonomous vehicle system . . . The process can be performed remotely and dynamically, for example, using an over-the-air update, without requiring the vehicle be brought to a service location”, where the second hardware processor is required to perform the “deploy[ment] to vehicles”, which transmits, “remotely and dynamically, for example, using an over-the-air update”, parameters that are set in the first machine learning model trained, “A new machine learning model is trained”, to the first hardware processor of the edge device, “then deployed to vehicles as an update to the autonomous vehicle system”); and . . . .
The remaining limitations are substantially the same as limitations of Claim 1, therefore it is rejected under the same rationale.
Regarding Claim 18, Karpathy teaches an information processing device in a machine learning system that comprises an edge device and the information processing device connected to the edge device via a network (Para. [0007], “FIG. 1B is a block diagram illustrating one embodiment of a system for generating training data” and FIG. 1B, where the machine learning system, as indicated by the inclusion of a “Deep Learning System 700” component, includes a vehicle “102” as a first device and a “Training Data Generating System 120” as a second device, which are connected via the “Network”; see also Para.[0023], “classifiers may be uploaded to a computer system within a vehicle, such that the classifier may be used to recognize specific image features or objects associated with the classifiers. The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system and used as training data for neural network systems”, where the “vehicle”, which as discussed above is “102”, is an edge device because it is at the edge of the “system”, distant from the “central server system”, and where the “server”, which is 120, see Para. [0080], “a computer server (e.g., the training data generation system 120)”, is an information processing device because it processes “training data”), and
selects a plurality of pieces of learning data for causing a first machine learning model to perform learning from among a plurality of pieces of input data . . . (Abstract, “Systems and methods for obtaining training data . . . includes . . . applying a neural network to the sensor data. A trigger classifier is applied to an intermediate result of the neural network to determine a classifier score for the sensor data. Based at least in part on the classifier score, a determination is made whether to transmit via a computer network at least a portion of the sensor data. Upon a positive determination, the sensor data is transmitted and used to generate training data”, where the “Systems” selects, “determination is made whether to transmit . . . [and use] to generate training data”, a plurality of pieces of learning data of learning data from a plurality of pieces of input data, where the selected “portion of the sensor data” are the learning data, selected from the input data of all “sensor data”; see also Para. [0081], “sensor data received is processed to create training data for training a machine learning model”, where the “training data” is used for causing a first “machine learning model” to perform learning through “training”; see also Para. [0080], “the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”, where the “server” information processing device, in combination other elements of the system, performs the selecting because, as a nonexclusive example, the “sensor data” “identified” by the classifier scores are used to “create training data”).
The remaining limitations are substantially the same as limitations of Claim 17, therefore it is rejected under the same rationale.
Regarding Claim 19, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein the first hardware processor generates the first evaluation value that becomes a larger value as the difference between the first input data and the k pieces of the candidate data is larger (Karpathy, Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Determine Trigger Classification Score 411” functionality, which generates the first evaluation value, “classification score”, see Karpathy, Para. [0077], “At 411, a trigger classifier score is determined”, which, in view of Kanno, becomes a larger value as the difference between the first input data and the k pieces of the candidate data is larger, see Kanno, Para. [0036], “ the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference. Here, it is assumed that the selecting unit 4 sorts the training data in ascending order based on a value indicating the difference”, where the calculated evaluation value, the rank used to “sort . . . in ascending order”, for selection of pieces that are greatly different from one another to allow for “obtaining examples of hard to find ‘edge cases’ of certain features” like “images of different tires on different roads in different driving environments”, see Karpathy, Para. [0023], “there may be a large number of vehicles being driven in disparate environments, which increases the likelihood of obtaining examples of hard to find ‘edge cases’ of certain features” and Karpathy, Para. [0037], “The training data may be enhanced by the inclusion of images of different tires on different roads. Additionally, the training data may be enhanced by images of different tires on different roads in different driving environments”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Teinemaa et al. (hereinafter Teinemaa) (“Temporal stability in predictive process monitoring”).
Regarding Claim 3, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the first evaluation value, the first hardware processor calculates, as the first evaluation value, a value (Karpathy, Fig. 4, where, as discussed above, the associated software and hardware required to execute the “Determine Trigger Classification Score 411” functionality are collectively performed by the first hardware processor, which, as also discussed above, is configured to calculate a first evaluation value, “classification score”, see Karpathy, Para. [0077], “At 411, a trigger classifier score is determined”) . . .
[according to a time-based analysis, involving] acquisition time of the first input data and acquisition time of second input data that is selected as one of the pieces of candidate data . . . [wherein, the] first input data among one or more pieces of the second input data different from the first input data (Karpathy, Para. [0021], “the sensor information may be captured in the normal course of operation of the vehicles. The sensor information may be used by the vehicles for certain automated driving features, such as lane navigation”, where the “sensor information” “captured in the normal course of operation” requires multiple stages of input data generation, including first input data and second input data, which must be different from the first to detect changing “features” for “automated driving”; see also Karpathy, Para. [0078], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415 . . . only sensor data with the highest score from the same location within the last 10 minutes may be retained as potential data”, where different first and second input data is also analyzed for multiple vehicles “from the same location”, which are selected as pieces of candidate data, “In the event the classifier score exceeds the threshold value, processing continues to 415”, is in part according to a time-based analysis of acquisition between the different input data, “within the last 10 minutes”).
Karpathy in view of Kanno do not explicitly disclose . . . according to a time difference between . . . immediately before the . . . .
However, Teinemaa teaches . . . [calculating a value] according to a time difference between . . . [the time of first data and the time of second data, where the second data is] immediately before the [first data] . . . . (Pg. 1308, Abstract, “Predictive process monitoring is concerned with the analysis of events produced during the execution of a business process . . . this paper defines a notion of temporal stability for binary classification tasks in predictive process monitoring”, where “analysis of events” requires data for a plurality of “events”, including first “event” data and second “event” data, where a value is calculated for the first event according to a time difference between it and the event immediately before it, “time since last event”, see Pg. 1318, “we apply some preprocessing on the raw datasets. In general, we use all the available case and event attributes without doing any feature extraction before encoding. Still, a few extra features are added to each event based on the timestamps, namely, hour, weekday, month, time since case start, and time since last event”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the calculation of a first evaluation value for use in selection of candidate data pieces as training data, where the value is calculated according to a time-based analysis of the acquisition times of first input data and different input data of Karpathy in view of Kanno with the calculation of a value according to a time difference between the time of first data and the time of second data, where the second data is immediately before the first data of Teinemaa in order to incorporate differences in acquisition time into the data selection process (Teinemaa, Pg. 1327, Para. 3, “Temporal stability characterizes how much successive prediction scores obtained for the same case (sequence of events) differ from each other”), which allows for data processing to remove volatility (Teinemaa, Pg. 1327, Para. 3, “For a temporally stable classifier, such successive prediction scores are similar to each other, resulting in a smooth time series, while in case of an unstable classifier, the resulting time series is volatile”; Teinemaa, Pg. 1308, Para. 1, “volatile predictions can mislead users of the system”) and allows for context-dependent decisions on whether input data should be incorporated into the training data (Karpathy, Para. [0078], “only sensor data with the highest score from the same location within the last 10 minutes may be retained as potential data”).
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Luo et al. (hereinafter Luo) (“Confident Learning: Estimating Uncertainty in Dataset Labels”).
Regarding Claim 8, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein the first machine learning model classifies input data into any of a plurality of classes (Karpathy, Para. [0056], “At 205, deep learning analysis of the sensor data is initiated . . . In various embodiments, the machine learning model is trained offline and installed onto the vehicle for performing inference on the sensor data. For example, the model may be trained to identify road lane lines, obstacles, pedestrians, moving vehicles, parked vehicles, drivable space, etc., as appropriate”, where the first machine learning model, “the machine learning model” used in “205, deep learning analysis”, classifies input data, “identif[ies]” “sensor data”, into a plurality of classes, “road lane lines, obstacles, pedestrians, moving vehicles, parked vehicles, drivable space, etc.,” for use in “autonomous driving”, see Karpathy, Para. [0060], “the results of the deep learning analysis are provided to vehicle control. For example, the results are used by a vehicle control module to control the vehicle for autonomous driving” and Karpathy, Para. [0057], “possible use cases may involve identifying: a curved road, an on ramp, an off ramp, the entrance to a tunnel, the exit of a tunnel, an obstacle in the road, a fork in the road, road lane lines or markers, drivable space, road signage, contents of signs (e.g., words, numbers, symbols, etc.), and/or other features as appropriate for autonomous driving”), and
in calculating of the second evaluation value (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which is output based on the configuration of the “machine learning model”, the second evaluation standard in this instance),
the second hardware processor acquires a classification probability of belonging to each of the classes (Karpathy, Fig. 5, where the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to receive “sensor data”, “the sensor data received at 501 includes data identified as potentially useful training data”, which includes the “classifier score”, see Karpathy, Para. [0029], “additional metadata is collected and retained along with the sensor data such as . . . the classifier score”, which is the probability of belonging to a class, see Karpathy, Para. [0057], “a higher classified score indicates a higher likelihood the sensor data is representative of the use case”; see also Karpathy, Para. [0023], “A multitude of these classifiers may be uploaded to a computer system within a vehicle, such that the classifier may be used to recognize specific image features or objects associated with the classifiers. The captured images that are designated by the classifier as including the particular feature or object can then be transmitted to a central server system and used as training data for neural network systems”, where a “multitude” of classifier scores identifying each of the classes, “captured images that are designated by the classifier as including the particular feature or object”, can be provided to the “server”)
obtained by inputting the first candidate data to the first machine learning model (Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the first candidate data, is part of the data input to the first machine learning model, “deep learning system”, which is analyzed by the “trigger classifier”, see Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”), and
calculates, as the second evaluation value, a value according to a degree of difference representing a difference between a classification probability of a class into which the first candidate data is classified as belonging among the classes and one or a plurality of classification probabilities for [comparison] . . . (Karpathy, Fig. 5, where, as discussed above, the associated software and hardware required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, is according to a degree of difference “a difference”, representing a difference between a classification probability of a class into which the first candidate data is classified among the classes of classification, “the selecting unit 4 determines a category to which each of training data belongs” and the classification probability of comparison, “correct answer data”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . one or a plurality of classes into which the first candidate data is classified as not belonging among the classes.
However, Luo teaches . . . [calculating a value according to a difference between a classification probability the first candidate data is classified into and] one or a plurality of classes into which the first candidate data is classified as not belonging among the classes (Pg. 594-595, Para. 4-1, “Suppose P(a) is the largest and P(b) is the second largest probability for example x, where a, b are class labels. “BT” tries to improve the P(a)−P(b). Intuitively, improving the value of P(a)−P(b) amounts to breaking the tie between P(a) and P(b), thus improving the classification confidence”, where “example x” is the first candidate data, which is classified into “class” “a” and not “class” “b”, which is one class “x” is not classified as belong to, and a value is calculated between the difference probabilities of the classes, “P(a)−P(b)”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the calculation, as the second evaluation value, a value according to a degree of difference representing a difference between a classification probability of a class into which the first candidate data is classified as belonging among the classes and one or a plurality of classification probabilities of comparison of Karpathy in view of Kanno with calculating a value according to a difference between a classification probability the first candidate data is classified into and for one or a plurality of classes into which the first candidate data is classified as not belonging among the classes of Luo in order to identify data associated with lower certainty model outputs (Luo, Pg. 595, Para. 1, “Intuitively, improving the value of P(a)−P(b) amounts to breaking the tie between P(a) and P(b), thus improving the classification confidence”, where a smaller “value of P(a)−P(b)” shows that the model’s “classification confidence” between “between P(a) and P(b)” is lower), which allows for removal of training data candidates with reduced utility for particular use cases (Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case” and Karpathy, Para. [0045], “Thus, the outside system 120 may receive sensor data 108 from a multitude of vehicles . . . For example, a portion of images transmitted to the system 120 may not include tires. In some embodiments, the entity may thus rapidly review and discard certain of the images. The remaining images may be aggregated into large training data sets and used to update the machine learning models executing on the vehicle”, where lower confidence scores, such as those incorrectly classified, “a portion of images transmitted to the system 120 may not include tires”, can be discarded based on their lower difference score).
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Mahajan et al. (hereinafter Mahajan) (“Towards Statistical Guarantees in Controlling Quality Tradeoffs for Approximate Acceleration”).
Regarding Claim 9, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the second evaluation value (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance),
the second hardware processor calculates, as the second evaluation value, a value according to a degree of difference . . . (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated representing a degree of difference between data, “correct answer data” and “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”)
[associated with] first evaluation data and second evaluation data the first hardware processor performs arithmetic processing with first arithmetic accuracy, according to the first machine learning model, the second hardware processor performs arithmetic processing . . . according to the first machine learning model (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where, as discussed above, the “vehicle” is the first information processing device, which executes the functionality of Fig. 4, and where the first evaluation data is all data produced during this “process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” because it data used to evaluate whether to “identify” data as “potential training data”; see also Karpathy, Para. [0039], “a deep learning system 700 of one or more processors, which is included in the vehicle 102”, where the first information processing device includes the first arithmetic processing device, “one or more processors”, used to perform arithmetic with first arithmetic accuracy, “complex . . . calculations”, see Karpathy, Para. [0106], “application-specific hardware or one or more physical computing devices (utilizing appropriate specialized executable instructions) may be necessary to perform the functionality, for example, due to the volume or complexity of the calculations involved or to provide results substantially in real-time”; see also Karpathy, Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, which as discussed above is the second information processing device, performs the functionality of “FIG. 5[’s] . . . flow diagram”, all the data generated during the “process” is the second evaluation data because it is used to evaluate “training data” and the “machine learning model” for “deploy[ment]”; see also Karpathy, Fig. 5 and Karpathy, Para. [0106], “application-specific hardware or one or more physical computing devices (utilizing appropriate specialized executable instructions) may be necessary to perform the functionality, for example, due to the volume or complexity of the calculations involved or to provide results substantially in real-time”, where the second arithmetic processing device, “hardware or one or more physical computing devices”, is “necessary to perform the functionality . . . [due to the] complexity of the calculations involved” in Fig. 5; see also Karpathy, Abstract, “Systems and methods for obtaining training data . . . includes . . . applying a neural network to the sensor data. A trigger classifier is applied to an intermediate result of the neural network to determine a classifier score for the sensor data. Based at least in part on the classifier score, a determination is made whether to transmit via a computer network at least a portion of the sensor data. Upon a positive determination, the sensor data is transmitted and used to generate training data” and Karpathy, Para. [0081], “sensor data received is processed to create training data for training a machine learning model”, where both arithmetic processing’s are according to the first machine learning model because they are in conformity with and dependent on “training . . . [the first] machine learning model”),
the first evaluation data (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where the first evaluation data is all data produced during “Fig. 4[‘s]” “process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” because it data used to evaluate whether to “identify” data as “potential training data”)
includes at least one of output data of the first machine learning model and intermediate data output from a first position inside the first machine learning model (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where, as discussed above, data generated as part of the “process for identifying potential training data” is first evaluation data, including both “output” “of the final layer” or “output” of an “intermediate” layer, either of which is a first position inside the first machine learning model, see Karpathy, Para. [0072] – [0076], “At 401, a deep learning analysis is initiated . . . with sensor data captured by sensors attached to a vehicle . . . At 403, inference using one layer of the deep learning analysis is completed. For example, a neural network . . . At 405, a determination is made whether the output of the layer analysis performed at 403 is a result of the final layer of the neural network. In the event the output is not the result of the final layer, for example, the output is an intermediate result . . . At 409, a determination is made whether the layer of the neural network and trigger conditions are appropriate for applying the trigger classifier”; see also Karpathy, Para. [0076], “For example, some use cases may be more efficient and produce high quality results using the intermediate result of a latter layer of the neural network. Other use cases may require an earlier intermediate result in order to identify useful examples of sensor data that meet the use case. In some cases, the trigger properties used to specify the conditions to apply the trigger classifier can be nested using multiple conditional checks and/or logical operators such as AND and OR operators”)
obtained by inputting the first candidate data to the first machine learning model (Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the first candidate data, is part of the data input to the first machine learning model, “deep learning system”, which is analyzed by the “trigger classifier”, see Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”), and
the second evaluation data (Karpathy, Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, performs the functionality of “FIG. 5[’s] . . . flow diagram”, and all the data generated during the “process” is the second evaluation data because it is used to evaluate “training data” and the “machine learning model” for “deploy[ment]”; see also Karpathy, Fig. 5 and Karpathy, Para. [0106], “application-specific hardware or one or more physical computing devices (utilizing appropriate specialized executable instructions) may be necessary to perform the functionality, for example, due to the volume or complexity of the calculations involved or to provide results substantially in real-time”, where, as discussed above, the second arithmetic processing device, “hardware or one or more physical computing devices”, is “necessary to perform the functionality . . . [due to the] complexity of the calculations involved” in Fig. 5)
includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data output from the first position inside the first machine learning model (Karpathy, Para. [0078] – [0080], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415 . . . At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”, where the second evaluation data includes “the identified sensor data” that is “transmitted to a computer server”, to perform the functionality of “FIG. 5[’s] . . . flow diagram”, see Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”; see also Karpathy, Para. [0077], “At 411, a trigger classifier score is determined. For example, a trigger classifier score is determined by applying the trigger classifier to the intermediate results of the neural network”, where the “identified sensor data”, which is “transmitted” to be part of the second evaluation data, corresponds to the output data/intermediate output data, “for example . . . intermediate results”, of the first position inside the first machine learning model, “the neural network”, because the “identified sensor data” is determined by a “classifier score” from “applying the trigger classifier to the intermediate results”; ”; see also Karpathy, Para. [0076], “For example, some use cases may be more efficient and produce high quality results using the intermediate result of a latter layer of the neural network. Other use cases may require an earlier intermediate result in order to identify useful examples of sensor data that meet the use case. In some cases, the trigger properties used to specify the conditions to apply the trigger classifier can be nested using multiple conditional checks and/or logical operators such as AND and OR operators”)
obtained by inputting the first candidate data to the first machine learning model (Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the first candidate data, is part of the data input to the first machine learning model, “deep learning system”, which is analyzed by the “trigger classifier”, so the first and second evaluation data are directly and indirectly, respectively, obtained from inputting the first candidate data to the first machine learning model, see Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . representing a difference between . . . with second arithmetic accuracy higher than the first arithmetic accuracy . . . .
However, Mahajan teaches . . . [performing an action when a value,] . . . representing a difference between . . . [data processed by a first arithmetic processing device and a second arithmetic processing device, is larger than another value] . . . (Pg. 2, Col. 2, Para. 2, “The difference between the imprecise accelerator output and the precise output is the accelerator error”; Pg. 4, Col. 1, Para. 5, “For each invocation, use the original precise result if the accelerator error exceeds the threshold”; see also Pg. 2, Col. 1, Para. 2, “The outputs from the accelerator are an approximation of the outputs that the core would have calculated”, where the “imprecise accelerator” is the first arithmetic processing device that produces the “imprecise” “output” and the “the core”, is the second arithmetic processing device that produces the “precise output”)
[wherein the second arithmetic processing device is configured as] with second arithmetic accuracy higher than the first arithmetic accuracy [of the first arithmetic processing device] . . . (Pg. 2, Col. 1, Para. 2, “Approximate accelerators trade small losses in output quality for significant performance and efficiency gains . . . When a processor core is augmented with an approximate accelerator, the core delegates the computation of frequently executed safe-to-approximate functions to the accelerator . . . The outputs from the accelerator are an approximation of the outputs that the core would have calculated”, where the “processor core” is an arithmetic processing device, which has higher arithmetic accuracy than the first arithmetic device, the “approximate accelerator” with outputs that are “an approximation of the outputs that the core would have calculated”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the machine learning system that calculates, as the second evaluation value, a value according to a degree of difference associated with the first and second evaluation data, wherein the system comprises the first hardware processor that performs arithmetic processing with first arithmetic accuracy, according to the first machine learning model, and the second hardware processor performs arithmetic processing with second arithmetic accuracy, according to the first machine learning model of Karpathy in view of Kanno with a second arithmetic processing device configured to execute information processing with higher arithmetic accuracy than a first arithmetic processing device, wherein an action is taken when a difference between a value from the data processed by the first arithmetic processing device and the data processed by the second arithmetic processing device is larger than a value of Mahajan in order to balance data processing tradeoffs of using either of the two information processing devices (compare Karpathy, Para. [0023], “The captured images that are designated by the classifier . . . can then be transmitted to a central server system and used as training data for neural network systems. Since the classifiers may leverage existing machine learning models already being executed by the vehicles in typical operation, the classifiers may be efficient in terms of processing requirements”, where “efficient” allocation of processing resources is used to handle “processing requirements” associated with actions taken on the “vehicles”, with Karpathy, Para. [0083], “at [the server] . . . a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the server must have the increased capacity to use “a highly accurate machine learning model . . . to confirm” the efficient calculations of the vehicles) by offloading model training to the second information processing device for training data instances associated with a significantly increased error rate on the first information processing device, while achieving energy and performance improvements by performing lower accuracy computations, where appropriate for the associated training data, using the first information processing device (Mahajan, Pg. 2, Col. 1, Para. 2-4, “When a processor core is augmented with an approximate accelerator, the core delegates the computation of . . . safe-to-approximate functions to the accelerator . . . Instead of executing the function, the core sends the function’s inputs to the accelerator and retrieves its outputs . . . MITHRA . . . provide[s] flexibility in controlling final quality loss and . . . maximize[s] the performance and energy benefits at any level of quality. MITHRA aims to only filter out those approximate accelerator invocations that cause relatively large quality degradation in the final output”).
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Yi et al. (hereinafter Yi) (“Transform consistency for learning with noisy labels”).
Regarding Claim 10, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the second evaluation value (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance),
the second hardware processor calculates, as the second evaluation value, a value according to a degree of difference representing a difference between . . . [data, to evaluate] . . . the first candidate data . . . [for selection of pieces of] . . . the first candidate data (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated representing a degree of difference between data, “correct answer data” and “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”),
the first evaluation data (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where the first evaluation data is all data produced during “Fig. 4[‘s]” “process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” because it data used to evaluate whether to “identify” data as “potential training data”)
includes at least one of output data of the first machine learning model and intermediate data output from a first position inside the first machine learning model (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where, as discussed above, data generated as part of the “process for identifying potential training data” is first evaluation data, including both “output” “of the final layer” or “output” of an “intermediate” layer, either of which is a first position inside the first machine learning model, see Karpathy, Para. [0072] – [0076], “At 401, a deep learning analysis is initiated . . . with sensor data captured by sensors attached to a vehicle . . . At 403, inference using one layer of the deep learning analysis is completed. For example, a neural network . . . At 405, a determination is made whether the output of the layer analysis performed at 403 is a result of the final layer of the neural network. In the event the output is not the result of the final layer, for example, the output is an intermediate result . . . At 409, a determination is made whether the layer of the neural network and trigger conditions are appropriate for applying the trigger classifier”; see also Karpathy, Para. [0076], “For example, some use cases may be more efficient and produce high quality results using the intermediate result of a latter layer of the neural network. Other use cases may require an earlier intermediate result in order to identify useful examples of sensor data that meet the use case. In some cases, the trigger properties used to specify the conditions to apply the trigger classifier can be nested using multiple conditional checks and/or logical operators such as AND and OR operators”), and
the second evaluation data (Karpathy, Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, performs the functionality of “FIG. 5[’s] . . . flow diagram”, and all the data generated during the “process” is the second evaluation data because it is used to evaluate “training data” and the “machine learning model” for “deploy[ment]”; see also Karpathy, Fig. 5 and Karpathy, Para. [0106], “application-specific hardware or one or more physical computing devices (utilizing appropriate specialized executable instructions) may be necessary to perform the functionality, for example, due to the volume or complexity of the calculations involved or to provide results substantially in real-time”, where, as discussed above, the second arithmetic processing device, “hardware or one or more physical computing devices”, is “necessary to perform the functionality . . . [due to the] complexity of the calculations involved” in Fig. 5)
includes data corresponding to the first evaluation data, which is any of the output data of the first machine learning model and the intermediate data output from the first position inside the first machine learning model (Karpathy, Para. [0078] – [0080], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415 . . . At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”, where the second evaluation data includes “the identified sensor data” that is “transmitted to a computer server”, to perform the functionality of “FIG. 5[’s] . . . flow diagram”, see Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”; see also Karpathy, Para. [0077], “At 411, a trigger classifier score is determined. For example, a trigger classifier score is determined by applying the trigger classifier to the intermediate results of the neural network”, where the “identified sensor data”, which is “transmitted” to be part of the second evaluation data, corresponds to the output data/intermediate output data from the first position inside the model, “for example . . . intermediate results”, of the first machine learning model, “the neural network”, because the “identified sensor data” is determined by a “classifier score” from “applying the trigger classifier to the intermediate results”; ”; see also Karpathy, Para. [0076], “For example, some use cases may be more efficient and produce high quality results using the intermediate result of a latter layer of the neural network. Other use cases may require an earlier intermediate result in order to identify useful examples of sensor data that meet the use case. In some cases, the trigger properties used to specify the conditions to apply the trigger classifier can be nested using multiple conditional checks and/or logical operators such as AND and OR operators”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . first evaluation data obtained by inputting . . . to the first machine learning model and second evaluation data obtained by inputting data that is obtained by partially changing . . . to the first machine learning model . . . .
However, Yi teaches . . . [calculating a value representing a degree of difference between a] first evaluation data obtained by inputting [input data] . . . to the first machine learning model and second evaluation data obtained by inputting data that is obtained by partially changing [the input data] . . . to the first machine learning model . . . (Pg. 1, Col. 2, Para. 2, “we feed the original and transformed (horizontally flip) images into one single network, and observe the Kullback-Leibler (KL) Divergence”, where the first evaluation data is the first output of the “one single network”, which is obtained by inputting, “feed[ing]”, “the original” data into the “one single network”, and where the second evaluation data is the second output of the “one single network”, which is obtained by inputting, “feed[ing]”, partially changing the input data, “transformed (horizontally flip) images” into the “one single network”, and where the “the Kullback-Leibler (KL) Divergence” is a value representing a degree of difference between the data).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the second hardware processor to calculate a second evaluation value for use, wherein the second evaluation value is a value according to a degree of difference representing a difference between data of Karpathy in view of Kanno with the calculating of a value representing a degree of difference between first evaluation data obtained by inputting input data to a first machine learning model and second evaluation data obtained by inputting data that is obtained by partially changing the input data to the first machine learning model of Yi in order to utilize a simple and effective method to identify potential training data as clean or noisy (Yi, Pg. 1, Col. 2, Para. 2, “we propose a simple and effective method to distinguish clean samples only using one single network. We find that the prediction consistency under different image transforms (such as scaling, rotation, flipping) in one network is beneficial to select clean samples”), which prevents poor performance of trained models, due to overfitting on noisy labels (Yi, Pg. 1, Para. 2, “Unfortunately, the obtained annotations inevitably contain noisy labels. As DNNs have the capability to memorize all training samples, they will eventually overfit the noisy labels, leading to poor generalization performance”).
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Liang et al. (hereinafter Liang) (“R-Drop: Regularized Dropout for Neural Networks”).
Regarding Claim 11, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the second evaluation vale (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance),
the second hardware processor calculates, as the second evaluation value, a value according to a degree of difference representing a difference between [data] and second evaluation data obtained by inputting the first candidate data to a second machine learning model (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated representing a degree of difference between data, “correct answer data”, and a plurality of pieces of second output data, “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”),
the first evaluation data (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where the first evaluation data is all data produced during “Fig. 4[‘s]” “process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” because it data used to evaluate whether to “identify” data as “potential training data”)
includes at least one of output data of the first machine learning model and intermediate data output from a first position inside the first machine learning model (Karpathy, Para. [0071]-[0072], “FIG. 4 is a flow diagram illustrating an embodiment of a process for identifying potential training data using . . . deep learning analysis of an autonomous driving system . . . [of] a vehicle” and Karpathy, Fig. 4, where, as discussed above, data generated as part of the “process for identifying potential training data” is first evaluation data, including both “output” “of the final layer” or “output” of an “intermediate” layer, either of which is a first position inside the first machine learning model, see Karpathy, Para. [0072] – [0076], “At 401, a deep learning analysis is initiated . . . with sensor data captured by sensors attached to a vehicle . . . At 403, inference using one layer of the deep learning analysis is completed. For example, a neural network . . . At 405, a determination is made whether the output of the layer analysis performed at 403 is a result of the final layer of the neural network. In the event the output is not the result of the final layer, for example, the output is an intermediate result . . . At 409, a determination is made whether the layer of the neural network and trigger conditions are appropriate for applying the trigger classifier”; see also Karpathy, Para. [0076], “For example, some use cases may be more efficient and produce high quality results using the intermediate result of a latter layer of the neural network. Other use cases may require an earlier intermediate result in order to identify useful examples of sensor data that meet the use case. In some cases, the trigger properties used to specify the conditions to apply the trigger classifier can be nested using multiple conditional checks and/or logical operators such as AND and OR operators”), and
the second evaluation data (Karpathy, Para. [0071], “The sensor data is then transmitted to a computer server and may be used to create training data for a revised machine learning model” and Karpathy, Para. [0081], “FIG. 5 is a flow diagram illustrating an embodiment of a process for creating training data”, where the “server”, performs the functionality of “FIG. 5[’s] . . . flow diagram”, and all the data generated during the “process” is the second evaluation data because it is used to evaluate “training data” and the “machine learning model” for “deploy[ment]”; see also Karpathy, Fig. 5 and Karpathy, Para. [0106], “application-specific hardware or one or more physical computing devices (utilizing appropriate specialized executable instructions) may be necessary to perform the functionality, for example, due to the volume or complexity of the calculations involved or to provide results substantially in real-time”, where, as discussed above, the second arithmetic processing device, “hardware or one or more physical computing devices”, is “necessary to perform the functionality . . . [due to the] complexity of the calculations involved” in Fig. 5)
includes data corresponding to the first evaluation data, which is any of the output data of the second machine learning model and the intermediate data output from the first position inside the second machine learning model (Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where, as discussed above, data generated during the process of Fig. 5 is the second evaluation data, which includes output data of the second machine learning model, output from the “highly trained machine learning model”, which is from the first position inside the second machine learning model, and which corresponds with the first evaluation data, because the “sensor data” “confirm[ed]” by the “highly trained learning model” is transmitted to the server based on the first evaluation data, “classifier score exceeds a threshold”, see Karpathy, Para. [0078] – [0080], “At 413, a determination is made whether the classifier score exceeds a threshold . . . In the event the classifier score exceeds the threshold value, processing continues to 415 . . . At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”, where the second evaluation data includes “the identified sensor data” that is “transmitted to a computer server”, to perform the functionality of “FIG. 5[’s] . . . flow diagram”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . the first evaluation data obtained by inputting the first candidate data to the first machine learning model . . . that is obtained by partially changing the first machine learning model . . . .
However, Liang teaches . . . [calculating a value representing a difference between] . . . the first evaluation data obtained by inputting the first candidate data to the first machine learning model . . . [and second evaluation data obtained by inputting the first candidate data to a second machine learning model] that is obtained by partially changing the first machine learning model . . . (Pg. 1, Para. 3, “each data sample goes through the forward pass twice, and each pass is processed by a different sub model”, where the first candidate data, “each data sample”, is “processed” “twice”, once by each of two “different sub model[s]”, where the first and second evaluation data are the outputs of the two “different sub model[s]”, the “two distributions of the model predictions”, see Pg. 3, Para. 4, “we can obtain two distributions of the model predictions, denoted as Pw1 (yi |xi) and Pw2 (yi |xi)”; see also Pg. 1, Abstract, “R-Drop minimizes the bidirectional KL-divergence between the output distributions of two sub models”, where a value of “KL-divergence” represents the difference between the first and second evaluation data, “KL-divergence between the output distributions of two sub models”; see also Pg. 3, Para. 4, “the two forward passes are indeed based on two different sub models (though in the same model)”, where the “two different sub models” are partially changed versions of each other, as versions of the “same model”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the use of the second hardware processor to calculate the second evaluation for training data selection, wherein the first evaluation value is obtained by inputting data into the first machine learning model and the second evaluation value is a value according to a degree of difference between data and second evaluation data obtained from inputting the first candidate data into a second machine learning model of Karpathy in view of Kanno with the calculation of a value representing a difference between first evaluation data obtained by inputting first candidate data to a first machine learning model and second evaluation data obtained by inputting the first candidate data to a second machine learning model that is obtained by partially changing the first machine learning model of Liang in order to select training data which produces consistent outputs across similar models (Liang, Pg. 1-2, Para. 3-1, “R-Drop forces the two distributions for the same data sample outputted by the two sub models to be consistent with each other, through minimizing the bidirectional Kullback-Leibler (KL) divergence between the two distributions”, where the output “distributions” will be “consistent” for similar “sub models” when “minimizing the bidirectional Kullback-Leibler (KL) divergence”), which will lead to more predictably consistent results during training and increased applicability of the trained models to inference use cases (Liang, Pg. 2, Para. 4, “We theoretically show that our R-Drop can reduce the inconsistency between training and inference of the dropout based models”, where, though framed in regard to reducing inconsistency by a model-altering mechanism, a sampling method to minimize KL divergence would also “reduce inconsistency”; see also Karpathy, Para. [0084] – [0086], “the training data of 503 is merged into existing training data sets. For example, an existing training data set applicable for most use cases is merged with the newly converted training data for improved coverage of a particular use case . . . At 509, the trained machine learning model is deployed. For example, the trained machine learning model is installed on a vehicle as an update for the autonomous learning system”, where when “merged with the newly converted training data” the training data would benefit from training data associated with predictably consistent results).
Claims 12-13 are rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Beluch et al. (hereinafter Beluch) (“The power of ensembles for active learning in image classification”).
Regarding Claim 12, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the second evaluation value (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance),
the hardware processor calculates, as the second evaluation value, a value [from] . . . a plurality of pieces of output data (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated from a plurality of pieces of output data, “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”), and
the pieces of output data are a plurality of inference results obtained by inputting the first candidate data to a . . . machine learning [model] . . . different from . . . the first machine learning model (Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where, as discussed above, the use of “a highly accurate machine learning model” for data evaluation requires the inputting of the candidate data “data identified as potentially useful” to the “machine learning model” , which as demonstrated by the “highly accurate” and its temporal location in the server “120”, is different from the first machine learning model in the “Deep Learning System 700”, see Karpathy, Fig. 1B, and the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated from a plurality of pieces of inference result output data, “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . representing variation among . . . plurality of . . . models . . . learned with learning parameters . . . learning parameters of . . . .
However, Beluch teaches . . . [calculating a value] representing variation among [a plurality of pieces of output data from a] . . . plurality of . . . models . . . learned with learning parameters [different from] learning parameters of [a reference model] (Pg. 9370, Col. 2, Para. 4, “the variance of
the softmax output vectors within the ensemble or within T forward passes can also be used as an acquisition function”, where “the variance” is a value representing variance of a plurality of pieces of output, “output vectors”, from “the ensemble”; see also Pg. 9372, Col.2, Para. 1, “an ensemble of five networks” and Pg. 9370, Col. 1, Para. 5, “all ensembles are trained with the same Dtrain and same network architecture, but different random weight initializations winit. One could also take additional measures to de-correlate the ensembles, such as bootstrapping or using different network architectures”, where “the ensemble” is a plurality of models, which will have different learning parameters than each other when “using different network architectures”, resulting in at least four of the “five networks” having different learning parameters than the learning parameters of another model of reference).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the calculation of an evaluation value from a plurality of pieces of output data for use in training data selection, wherein the plurality of pieces of output data are obtained by inputting the first candidate data to a machine learning model different from the first machine learning model of Karpathy in view of Kanno with the calculation of a value representing variation among a plurality of pieces of output data from a plurality of models learned with learning parameters different from learning parameters of a reference model of Beluch in order to select desirable training data (Beluch, Pg. 9368, Col. 2, Para. 1, “Starting with an initial (small) data set to train a model, new data-points to be labeled (e.g. by a human expert) are selected with a so-called acquisition function. This function ranks unlabeled data by “how desirable” label information is expected to be for each data-point. Commonly used acquisition functions are based on criteria such as variance reduction”; see also Beluch, Pg. 9370, Col. 1, Para. 5, “all ensembles are trained with the same Dtrain and same network architecture, but different random weight initializations winit. One could also take additional measures to de-correlate the ensembles, such as bootstrapping or using different network architectures”, where “de-correlat[ing]” the models would further allow for an assessment of variance) using a method with demonstrated state-of-the-art performance at training data acquisition (Beluch, Pg. 9375, Col. 2, Para. 2, “We compare the performance of acquisition functions and uncertainty estimation methods for active learning with CNNs on image classification tasks. We show that ensemble-based uncertainties consistently outperform other methods of uncertainty estimation (in particular MC Dropout) and lead to state-of-the-art active learning performance”).
Regarding Claim 13, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein in the calculating of the second evaluation value (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance),
the second hardware processor calculates, as the second evaluation value, a value based on a degree of difference representing a difference between . . . [data] and each of one or more pieces of second output data (Karpathy, Fig. 5, where, as discussed above, the second hardware processor is required to execute the “Convert Sensor Data Into Training Data 503” functionality, which is configured to evaluate whether each of the pieces of candidate data, “the sensor data received at 501 includes data identified as potentially useful training data” is effective when being used for learning of the first machine learning model, “confirm whether the sensor data represents the targeted use case”, by “a highly accurate machine learning model”, see Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where the use of “a highly accurate machine learning model” for data evaluation requires the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated representing a degree of difference between data, “correct answer data”, and a plurality of pieces of second output data, “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”),
the first output data is an inference result obtained by inputting the first candidate data to the first machine learning model (Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the first candidate data, is part of the data input to the first machine learning model, “deep learning system”, which is analyzed by the “trigger classifier”, see Karpathy, Para. [0071], “The trigger classifier analyzes sensor data at least partially analyzed by the deep learning system to identify whether the sensor data meets particular use cases that warrant retaining the sensor data”, where the “trigger classifier result” is the first output data), and
the one or more pieces of second output data are respectively one or more inference results obtained by inputting the first candidate data to one or more machine learning models . . . different from . . . the first machine learning model (Karpathy, Para. [0083], “At 503, . . . the sensor data received at 501 includes data identified as potentially useful training data . . . In some embodiments, the data is reviewed to determine whether the sensor data accurately represents the target use case . . . For example, a highly accurate machine learning model is used to confirm whether the sensor data represents the targeted use case”, where, as discussed above, the use of “a highly accurate machine learning model” for data evaluation requires the inputting of the candidate data “data identified as potentially useful” to the “machine learning model” , which as demonstrated by the “highly accurate” and its temporal location in the server “120”, is different from the first machine learning model in the “Deep Learning System 700”, see Karpathy, Fig. 1B, and the output of an evaluation value, the second evaluation value in this instance, which, in view of Kanno, the value, “a value indicating the difference”, is calculated from a plurality of pieces of inference result second output data, “a category to which each of training data belongs”, see Kanno, Para. [0036], “the selecting unit 4 determines a category to which each of training data belongs by applying each of the training data to the first model . . . the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data . . . a value indicating the difference”).
The reasons for obviousness, in regard to the combination of Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 1 above and remain applicable here.
Karpathy in view of Kanno do not explicitly disclose . . . first output data . . . learned with learning parameters . . . learning parameters of . . . .
However, Beluch teaches . . . [calculating a value based on a degree of difference between] first output data . . . [and second output data from a machine learning model] learned with learning parameters [different from] learning parameters of [the model used to generate the first output data] (Pg. 9370, Col. 2, Para. 4, “the variance of the softmax output vectors within the ensemble or within T forward passes can also be used as an acquisition function”, where “the variance of” is a value based on a degree of difference between two sets of output data, “output vectors”, output from an “the ensemble”; Pg. 9370, Col. 1, Para. 5, “all ensembles are trained with the same Dtrain and same network architecture, but different random weight initializations winit. One could also take additional measures to de-correlate the ensembles, such as bootstrapping or using different network architectures”, where “the ensemble” is a plurality of models, necessarily requiring a first and second model associated with first and second output data, which will have different learning parameters than each other when “using different network architectures”).
The reasons for obviousness, in regard to the combination of Beluch with Karpathy in view of Kanno, were discussed in regard to the rejection of Claim 12 above and remain applicable here.
Claim 14 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno, Pemantle (“A survey of random processes with reinforcement”) and Derakhshan et al. (hereinafter Derakhshan) (“Continuous Deployment of Machine Learning Pipelines”).
Regarding Claim 14, Karpathy in view of Kanno teach the machine learning system according to claim 1, wherein the second hardware processor further transmits . . . information . . . [for selection of data] as the learning data, to the first hardware processor of the information processing device . . . (Karpathy, Para. [0048] – [0049], “The system 120 may then transmit information to a portion of vehicles to execute the same classifier when proximate to the particular real-world area. In this way, the system 120 may ensure that it is able to obtain a greater quantity of training data based on this same sensor. Furthermore, the system 120 may instruct vehicles to transmit sensor data even if the above-described classifier does not assign a classifier score greater than a threshold”, where the second hardware processor in “The system 120”, transmits information for selection of data as the learning data, such as “information to . . . execute the same classifier when proximate to the particular real-world area”, to the first information processing device, the “vehicle”, which requires the associated first hardware processor that is configured for the functionality), and
the first hardware processor receives the . . . information, and makes a probability of selecting, as the candidate data, the input data acquired in a first . . . range that is after a first time being a time of the input data indicated by the . . . information and that is determined with the first time as a standard, to be higher than a probability of selecting, as the candidate data, the input data acquired in a . . . range other than the first . . . range (Karpathy, Para. [0046] – [0049], “each classifier may execute for a particular period of time before being swapped for another classifier . . . The system 120 may then transmit information to a portion of vehicles to execute the same classifier when proximate to the particular real-world area. In this way, the system 120 may ensure that it is able to obtain a greater quantity of training data based on this same sensor. Furthermore, the system 120 may instruct vehicles to transmit sensor data even if the above-described classifier does not assign a classifier score greater than a threshold . . . In this example, the system 120 may instruct any vehicle within a threshold distance of that real-world location to transmit sensor data (e.g., images) even if their classifiers do not generate a classifier score greater than a threshold . . . In this way, the outside system 120 may override the classifier and cause the particular vehicle to transmit sensor data”, where the “transmit[ted] information”, which is received by the first hardware processor inside the “vehicles”, causes the probability of selecting input data as candidate data to be higher for input data collected after transmission that is within a first range, “ensure that it is able to obtain a greater quantity of training data . . . [by] overrid[ing] the classifier”, as compared with another range as candidate data, “any vehicle” outside of the “threshold distance of that real-world location”, and which is determined with the first time as a standard, “each classifier may execute for a particular period of time . . . The system 120 may then transmit information to a portion of vehicles to execute the same classifier . . . [and] the system 120 may instruct vehicles to transmit sensor data even if the above-described classifier does not assign a classifier score greater than a threshold”, such that, as compared with the standard of the first time, the “threshold” is comparatively lowered, and where the range must be determined in advance for it to be indicated by the transmitted information, “the system 120 may instruct”, such as to be after a first time being a time of the input data, “each classifier may execute for a particular period of time before being swapped for another classifier”).
Karpathy in view of Kanno do not explicitly disclose . . . employment . . . indicating that corresponding input data is selected . . . each time the learning data is selected . . . employment . . . time . . . employment . . . time . . . time . . . .
However, Pemantle teaches . . . [information comprising] . . . employment . . . indicating that corresponding input data is selected . . . each time the learning data is selected . . . [wherein the] . . . employment [increases the probability that data in the range of the data associated with the] . . . employment [will be selected] (Pg. 4, Para. 5, “The original Polya urn model . . . has an urn that begins with one red ball and one black ball. At each time step, a ball is chosen at random and put back in the urn along with one extra ball of the color drawn, this process being repeated”, where “At each step . . . repeated”, information is provided comprising employment indicating that corresponding data is selected, “a ball is chosen at random and put back in the urn along with one extra ball of the color drawn”; see also Pg. 4, Para. 2, “the probability of choosing a ball of a given type is equal to the proportion of that type in the urn”, where the employment, “choosing a ball”, increases the number of data in that range, “of the [same] color drawn” in this instance, which increases the “proportion of that type” and therefore “the probability of choosing”; see generally Pg. 25, Para. 2, “Urn models: applications . . . use reinforcement models (mostly urn models) . . . to provide quick and robust algorithms” and Pg. 35-36, Section “4.4 Learning”, where “urn model[s]” can be used for data sampling algorithms and in learning data applications).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the transmission of information for selection of data as learning data from the second hardware processor to the first hardware processor of the information processing device, wherein the information makes a probability of selecting, as the candidate data, the input data acquired in a first range that is after a first time being a time of the input data indicated by the information and that is determined with the first time as a standard, to be higher than a probability of selecting, as the candidate data, the input data acquired in a range other than the first range of Karpathy in view of Kanno with the information comprising employment indicating that corresponding input data is selected each time the learning data is selected, wherein the employment increases the probability that data in the range associated with the employment will be selected of Pemantle in order to utilize a quick and robust algorithm for data selection (Pemantle, Pg. 25, Para. 2, “Urn models: applications . . . use reinforcement models (mostly urn models) . . . to provide quick and robust algorithms”), which has achieved rigorously tested results across multiple domain applications (Pemantle, Pg. 43, Para. 4, “models now abound in a variety of social science disciplines, including psychology, sociology [BL03], public health [EL04], political science [OMH+04]. The discussion here will concentrate on a few game-theoretic applications in which rigorous results have been obtained”) in order to prioritize selection of training data for specific use cases, by incorporating reinforcement models to select data with similar attributes (Karpathy, Para. [0004], “Typically, the performance of the deep learning system is limited at least in part by the quality of the training set used to train the model. In many instances, significant resources are invested in collecting, curating, and annotating the training data. The effort required to create the training set can be significant and is often tedious. Moreover, it is often difficult to collect data for particular use cases that a machine learning model needs improvement on”).
Additionally, Derakhshan teaches . . . [selecting training data for a first] time [range] . . . [at a higher probability compared with another] time [range other than the first] time [range] (Pg. 6, Col. 2, Para. 3, “The time-based sampling strategy assigns weights to every data chunk based on their timestamp such that recent chunks have a higher probability of being sampled. The window-based sampling strategy is similar to the uniform sampling, but instead of sampling from the entire historical data, the data manager samples the data from a given time range”, where “sampling” assigns a “higher probability” for some “time range” “chunk[s]” as compared to other “chunk[s]”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the selection of input data as candidate data, wherein, after each selection, feedback information is transmitted from the second hardware processor to the first hardware processor so that input data in a range of the selected data will have a higher probability of being selected than input data in another range of Karpathy in view of Kanno and Pemantle with the selecting training data for a first time range at a higher probability compared with another time range other than the first time range of Derakhshan in order to prioritize the selection of data from a specific time range, in instances where a machine learning model will have better performance when applied to a particular use case, if trained on data from a specific time range, such as data from vehicles traveling near a known, but impermanent, highway hazard (Derakhshan, Pg. 6, Col. 2, Para. 3, “Based on the specific use-case, the user chooses the appropriate sampling strategy. In many real-world use cases (e.g., e-commerce and online advertising), the deployed model should adapt to the more recent data. Therefore, the time-based and window-based sampling provide more appropriate samples for training”; see also Karpathy, Para. [0046] – [0049], “each classifier may execute for a particular period of time before being swapped for another classifier . . . The system 120 may then transmit information to a portion of vehicles to execute the same classifier when proximate to the particular real-world area. In this way, the system 120 may ensure that it is able to obtain a greater quantity of training data based on this same sensor. Furthermore, the system 120 may instruct vehicles to transmit sensor data even if the above-described classifier does not assign a classifier score greater than a threshold . . . In this example, the system 120 may instruct any vehicle within a threshold distance of that real-world location to transmit sensor data (e.g., images) even if their classifiers do not generate a classifier score greater than a threshold . . . In this way, the outside system 120 may override the classifier and cause the particular vehicle to transmit sensor data”).
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Karpathy in view of Kanno and Pradeep et al. (hereinafter Pradeep) (Patent Pub. No. US 2019/0220698 A1).
Regarding Claim 20, Karpathy in view of Kanno teach the machine learning system according to claim 19, wherein the first hardware processor calculates, as the first evaluation value . . . when each of the plurality of pieces of input data is image data (Karpathy, Fig. 4, where, as discussed above, the first hardware processor is required to execute the “Determine Trigger Classification Score 411” functionality, which is configured generates the first evaluation value, “classification score”, see Karpathy, Para. [0077], “At 411, a trigger classifier score is determined”, when each of the plurality of pieces of input data is image data, see Karpathy, Pg. [0024], “his specification describes vehicles obtaining sensor information, such as images” and Karpathy, Para. [0060], “Para. [0060], “the intermediate results of the deep learning analysis at 205 are utilized for identifying training data at 207 and transmitting the identified sensor data at 209”, where the plurality of pieces of “sensor data” is “image[]” data).
Karpathy in view of Kanno do not explicitly disclose . . . a Sum of Absolute Difference (SAD) or a Sum of Squared Difference (SSD) . . . .
However, Pradeep teaches . . . [analyzing training data using] a Sum of Absolute Difference (SAD) or a Sum of Squared Difference (SSD) . . . (Para. [0046], “the training data generator 238 is configured (for example, by the device controller interface 220) to apply various criteria to select training event instances for which corresponding device-generated training data 240 is generated. One such criteria may be a similarity of training data provided by the training data selector 256 for a current training event instance . . . techniques such as . . . sum of squared difference may be used”).
Before the effective filing date of the invention, it would have been obvious to one of ordinary skill in the art to combine the first hardware processor that calculates the first evaluation value when each of the plurality of pieces of input data is image data of Karpathy in view of Kanno with the analysis of training data using a Sum of Absolute Difference (SAD) or a Sum of Squared Difference (SSD) of Pradeep in order utilize a Sum of Squared Difference as a value for candidate data selection (Karpathy, Para. [0080], “At 415, the identified sensor data is transmitted. For example, the sensor data identified is transmitted to a computer server (e.g., the training data generation system 120) where it may be used to create training data”; Kanno, Para. [0036], “the selecting unit 4 calculates a difference between a category determined for training data and correct answer data corresponding to the training data, and further sorts the training data based on the difference”) such as to analyze the image data using a well-established measurement to ensure selected pieces satisfy training data criteria (Pradeep, Para. [0046], “the training data generator 238 is configured (for example, by the device controller interface 220) to apply various criteria to select training event instances for which corresponding device-generated training data 240 is generated. One such criteria may be a similarity of training data provided by the training data selector 256 for a current training event instance to training data for a previous training event instance. For example, techniques such as . . . sum of squared difference may be used to determine the training data for the current event instance is too similar to previous training data to be likely to meaningfully improve upon previous device-generated training data 240”), which allows for greater specificity of model training and, as a result, increased inference accuracy of trained models on specific use cases (Karpathy, Para. [0023], “there may be a large number of vehicles being driven in disparate environments, which increases the likelihood of obtaining examples of hard to find ‘edge cases’ of certain features” and Karpathy, Para. [0037], “The training data may be enhanced by the inclusion of images of different tires on different roads. Additionally, the training data may be enhanced by images of different tires on different roads in different driving environments”).
Response to Arguments
Applicant's arguments filed on April 20th, 2026 have been fully considered. Each argument is addressed in detail below.
I. Applicant argues the claims, as amended, do not invoke 35 USC § 112(f) (Applicant’s Remarks, 04/20/2026, Pg. 22, Section “III. Claim Interpretation under 35 U.S.C. § 112(f) and Rejection under 35 U.S.C. 112(b)”).
The claims, as amended, do not invoke 35 USC § 112(f). As a result, the claims are not interpreted under 35 USC § 112(f).
II. Applicant argues the rejections of the claims under 35 USC § 112 should be withdrawn (Applicant’s Remarks, 04/20/2026, Pg. 22, Section “III. Claim Interpretation under 35 U.S.C. § 112(f) and Rejection under 35 U.S.C. 112(b)”).
Applicant’s amendments to the claims have corrected most, but not all, of the issues identified as grounds for rejection under 35 USC § 112 in the January 6th, 2026 Office Action. Unresolved issues are identified and discussed at length above. Additionally, as also identified and discussed at length above, Applicant’s amendments introduced additional indefiniteness, which necessitated new grounds of rejection under 35 USC § 112.
III. Applicant argues the objections to the claims should be withdrawn (Applicant’s Remarks, 04/20/2026, Pg. 22, Section “IV. Claim Objections”).
Applicant’s amendments to the claims have overcome each and every objection to the claims, as previously set forth in the January 6th, 2026 Office Action. As a result, these objections have been withdrawn.
IV. Applicant argues the rejections of the claims under 35 USC § 101 should be withdrawn (Applicant’s Remarks, 04/20/2026, Pg. 23-33, Section “V. Rejection under 35 U.S.C. § 101”).
1) First, Applicant argues the claims are eligible under Step 2A, Prong 1 because the claims, as amended, recite “generation of time-series input data based on observed surroundings, execution of inference processing using a trained machine learning model, calculation of evaluation values based on measured differences among sequential input data, transmission of candidate data over a network, and analysis of inference results or intermediate results performed by hardware processors. These operations are concrete, hardware-anchored operations performed by distinct hardware processors ("first hardware processor" and "second hardware processor") of distributed information processing devices ("first information processing device" and "second information processing device")”, which “cannot be practically performed in the human mind” (Applicant’s Remarks, Pg. 23-27).
However, claims that require a computer may still recite a mental process (see MPEP 2106.04(a)(2)(III)(C)).
As a result, the argument is not persuasive.
2) Second, Applicant argues the claims are eligible under Step 2A, Prong 2 because the specification discloses “selectively relearning the machine learning model using a plurality of pieces of input data that exhibit large variation, while partitioning processing load between devices . . . [and] selecting input data that are greatly different from one another as effective learning data, because relearning using such highly varied data enables the machine learning model to accurately perform inference over a wide range of input data” as a technical solution to the problem of “excessive arithmetic burden on edge devices . . . require[ing] transmission of large volumes of input data to a cloud device, thereby increasing communication cost and reducing efficiency . . . [and] difficulty selecting effective learning data and that learning using highly similar input data does not yield accurate inference over extensive input data” (Applicant’s Remarks, Pg. 27-28). Additionally, Applicant argues the claims, as amended, reflect the improvement described in the specification by reciting “the first hardware processor performs ... calculating a value according to a degree of a difference between the first input data and k pieces of the candidate data immediately before the first input data ... such that pieces of input data that are greatly different from one another are selected as the pieces of the candidate data” and that “the second hardware processor performs ... calculating the second evaluation value ... by analyzing an inference result or an intermediate result obtained by inputting the first candidate data to a machine learning model”, which “provides specific technical solutions to problems in the field of artificial intelligence and machine learning” that integrate the abstract ideas into a practical application (Applicant’s Remarks, Pg. 28-31).
However, while the specification sets forth the alleged improvements to athematic processing on edge computing devices and machine learning training data selection, these improvements are asserted in a conclusory manner and without the detail necessary to be apparent to a person of ordinary skill in the art as to how the disclosure subject matter variation-based data selection and partitioning results in the asserted improvements (see MPEP 2106.04(d)(1)).
Additionally, the claims recite limitations, such as variation-based data selection and evaluation based on calculated results, that have broad applicability across many fields of endeavor (see MPEP 2106.05(f)), such as to fail to reflect the alleged improvements (see MPEP 2106.04(d)(1)).
As a result, the argument is not persuasive.
3) Third, Applicant argues the claims are eligible under Step 2B because the asserted improvements allow for “select[ion of] highly varied input data for relearning while distributing lightweight and heavyweight processing between resource-limited and resource-rich devices, thereby reducing data transmission and improving inference accuracy over a wide range of input data” such as to constitute a “specific technique designed to enable efficient relearning using highly varied input data while conserving computational and communication resources”, which amounts to significantly more than any alleged abstract idea (Applicant’s Remarks, Pg. 31-33).
However, as discussed above, the asserted improvements are recited in the specification in a conclusory manner (see MPEP 2106.04(d)(1)) and the claims recite limitations with broad applicability across many fields of endeavor (see MPEP 2106.05(f)), such as to fail to reflect the alleged improvements (see MPEP 2106.04(d)(1)).
As a result, the argument is not persuasive.
IV. Applicant argues the rejections of the claims under 35 USC § 102 and 35 USC § 103 should be withdrawn (Applicant’s Remarks, 04/20/2026, Pg. 33-36, Sections “VI. Rejections under 35 U.S.C. §§ 102 and 103” and “VII. New claims”).
In response to Applicant’s amendments, the previously communicated rejections under 35
U.S.C. § 102 and 35 U.S.C. § 103, have been withdrawn. However, Applicants arguments are not
persuasive in light of the new grounds for rejection, under 35 U.S.C. § 103, discussed in detail above. The new grounds of rejection rely on new combinations of the existing prior art of record and new prior art of record to teach the new combinations of elements in the amended claims, which were not presented in these arrangements in any of the previously presented claims. As a result, Applicant arguments against the previously communicated rejections under 35 U.S.C. § 102 and 35 U.S.C. § 103 are rendered moot.
However, in order to expedite prosecution and in the interest of clarity, any arguments still relevant to the new grounds of rejection are discussed below.
Specifically, Applicant argues Karpathy in view of Kanno would be insufficient to fully teach each and every element of the amended independent claims because “Kanno discloses that training data having a small difference from correct answer data is appropriate as training data ([0038]). Thus, Kanno would have motivated a person of ordinary skill to select input data having small differences, not "input data that are greatly different from one another," as recited in amended claim 1. Therefore, Kanno fails to cure the deficiencies of Karpathy to teach or suggest, at least, calculating a value representing a degree of a difference between the first input data and k pieces of the candidate data immediately before the first input data among the pieces of candidate data, k being an integral number equal to or larger than 1, such that pieces of input data that are greatly different from one another are selected as the candidate data; and as recited in amended independent claim 1” (Applicant’s Remarks, Pg. 16) (internal quotation marks omitted).
However, as discussed in detail above, a person of ordinary skill in the art would have been motivated by Karpathy in view of Kanno to modify the systems and methods disclosed in Karpathy to arrive at the claimed subject matter of the independent claims (see MPEP 2143(I)). Additionally, Applicant’s arguments against Kanno amount to attacking references individually where the rejections are based on combinations of references (see MPEP 2145 (IV)). Specifically, the rejection is based on a combination of Karpathy in view of Kanno. However, Applicant argues Karpathy fails to disclose elements that Kanno is relied upon to teach and argues Kanno fails to disclose elements that Karpathy is relied upon to teach. Furthermore, while Applicant points out that Kanno discloses its methodology is appropriate for selection of training data based on small differences, these are small differences between “correct answer data” and the “training data”, not small differences among pieces of training data (Kanno, Para. [0038], “It can be said that training data having a small difference from correct answer data is appropriate as training data used for learning the first model. In addition, it can be said that training data having a large difference from correct answer data is inappropriate as training data used for learning the first model”). As a result, Kanno does not teach away from the proposed combination because it does not address, and therefore certainly does not discourage, the proposed combination of selection of training data, such that pieces of input data that are greatly different from one another are selected as the candidate data (see MPEP 2144.05(III)(B)).
As a result, the argument is not persuasive.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MATTHEW BRYCE GOLAN whose telephone number is (571)272-5159. The examiner can normally be reached Monday through Friday, 8:00 AM to 5:00 PM ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached at (571) 270-3428. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/MATTHEW BRYCE GOLAN/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights