DETAILED ACTION
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination
2. A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 08 December 2025 [hereinafter Response] has been entered, where:
Claims 1, 2, 7, 8, 9, 14, 16, 19, and 20 have been amended.
Claims 10-13 and 15 have been cancelled.
Claims 1-9, 14, and 16-20 are pending.
Claims 1-9, 14, and 16-20 are rejected.
Claim Rejections - 35 U.S.C. § 101
3. 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.
4. Claims 9, 14, and 16 are rejected under 35 U.S.C. § 101 because the claimed invention is directed to an abstract idea without significantly more.
Claim 9 depends directly or indirectly from claim 1. Claim 9 recites a method, which is a process, and thus one of the statutory categories of patentable subject matter. (35 U.S.C. § 101).
However, under Step 2A Prong One, the claim recites the limitations of “[(f)] generating, by each encoding layer of the plurality of encoding layers of the second machine learning model, a layer output based at least in part on the corresponding intermediate output of the intermediate encoding layer of the first machine learning model.” The activity of “[(f)] generating,” in the nature of a prediction activity, can practically be performed in the human mind, including, for example, observations, evaluations, judgments, and opinions, and accordingly is a mental processes, (MPEP § 2106.04(a)(2) sub III), which is one of the groupings of abstract ideas. Thus, claim 9 recites an abstract idea.
Under Step 2A Prong Two, the claim as a whole is not integrated into a practical application, because the additional elements recited in the claim beyond the identified judicial exception include a “graphic processing unit,” a “temporary store,” and a “central processing unit,” which are recited in claim 1. These additional elements are generic computer components used to implement the abstract idea, and do not serve to integrate the abstract idea into a practical application. (MPEP § 2106.05(f)). The claim also includes additional elements of a “first machine learning model” and a “second machine learning model” recited in claim 1, which are recited at such a high level of generality as to be generic computer components used to implement the abstract idea, and do not serve to integrate the abstract idea into a practical application. (MPEP § 2106.05(f)).
The claim also recites additional elements of “[(a)] receiving data for use in generation of a machine learning model output,” “[(b)] ingesting the data with a first machine learning model,” “[(c)] receiving the at least one intermediate output from the temporary store,” “[(d)] ingesting the at least one intermediate output from the temporary store by another intermediate encoding layer in a second machine learning model,” and “[(e))] outputting a prediction with the second machine learning model,” which are recited in claim 1. The additional elements of “[(a)] receiving,” “[(b), (d)] ingesting,” and “[(f)] outputting” are pre-processing insignificant extra-solution activities of mere data gathering and transfer, (MPEP § 2106.05(g)), which does not integrate the abstract idea into a practical application.
Also, the independent claim from which the instant claim depends recites more details or specifics to the additional element of “[(c)] receiving at least one intermediate output,” where “[(c.1)] the at least one intermediate output comprising an output of an intermediate encoding layer in the first machine learning model,” and accordingly, is merely more specific to the additional element. The claim also recites more details or specifics to the additional element of “[(c)] receiving at least one intermediate output, where “[(c.1)] the at least one intermediate output comprises a plurality of intermediate outputs, [(c.1.1)] each of the plurality of intermediate outputs being generated by a unique encoding layer of the first machine learning model” of claim 7, which depends from claim 1, and accordingly, is merely more specific to the additional element. The claim also recites “[(b)] . . . a first machine learning model executed on a Graphic Processing Unit (“GPU”),” and “[(d.1)] wherein the second machine learning model is executed on a Central Processing Unit (“CPU”).” Generally linking the abstract idea to a field of use (that is, specifying the intended execution of models on general hardware devices) does not provide an inventive concept. (MPEP § 2106.05(h)). Also, execution on generic computer components (GPU, CPU) does not serve to integrate the abstract idea into a practical application. Therefore, claim 9 is directed to the abstract idea.
Finally, under Step 2B, the additional elements, taken alone or in combination, do not represent significantly more than the abstract idea itself. The additional elements include a “graphic processing unit,” a “temporary store,” and a “central processing unit,” which are recited in claim 1. These additional elements are generic computer components used to implement the abstract idea, and do not amount to significantly more than the abstract idea. (MPEP § 2106.05(f)). The claim also includes additional elements of a “first machine learning model” and a “second machine learning model” recited in claim 1, which are recited at such a high level of generality as to be generic computer components used to implement the abstract idea, and do not amount to significantly more than the abstract idea. (MPEP § 2106.05(f)).
The claim also recites additional elements of “[(a)] receiving data for use in generation of a machine learning model output,” “[(b)] ingesting the data with a first machine learning model,” “[(c)] receiving the at least one intermediate output from the temporary store,” “[(d)] ingesting the at least one intermediate output from the temporary store by another intermediate encoding layer in a second machine learning model,” and “[(e)] outputting a prediction with the second machine learning model,” which are recited in claim 1. The additional elements of “[(a)] receiving,” “[(b), (d)] ingesting,” and “[(f)] outputting” are well-understood, routine, and conventional activities of storing and retrieving information in memory, (MPEP § 2016.05(d) sub II.iv), which does not amount to significantly more than the abstract idea.
Also, the independent claim from which the instant claim depends recites more details or specifics to the additional element of “[(c)] receiving at least one intermediate output,” where “[(c.1)] the intermediate output comprising an output of an intermediate layer in the first machine learning model,” and accordingly, is merely more specific to the additional element. The claim also recites more details or specifics to the additional element of “[(c)] receiving at least one intermediate output, where “[(c.1)] the at least one intermediate output comprises a plurality of intermediate outputs, [(c.1.1)] each of the plurality of intermediate outputs generated by a unique layer of the first machine learning model” of claim 7, which depends from claim 1, and accordingly, is merely more specific to the additional element. The claim also recites “[(b)] . . . a first machine learning model executed on a Graphic Processing Unit (“GPU”),” and “[(d.1)] wherein the second machine learning model is executed on a Central Processing Unit (“CPU”).” Generally linking the abstract idea to a field of use (that is, specifying the intended execution of models on general hardware devices) does not provide an inventive concept. (MPEP § 2106.05(h)). Also, execution on generic computer components (GPU, CPU) does not amount to significantly more than the abstract idea. Therefore, claim 9 is subject-matter ineligible.
Claim 14 depends directly or indirectly from claim 1. The claim further recites the limitation of “[(f)] generating, by an encoding layer of the plurality of encoding layers of the second machine learning model, the layer output based at least in part on one or more outputs generated by one or more preceding encoding layers of the second machine learning model.” The activity of “[(f)] generating,” in the nature of a prediction activity, can practically be performed in the human mind, including, for example, observations, evaluations, judgments, and opinions, and accordingly is a mental processes, (MPEP § 2106.04(a)(2) sub III), which is one of the groupings of abstract ideas. The additional elements of the claim does not serve to integrate the abstract idea into integrated into a practical application, (see MPEP § 2106.04(d)), nor do the additional elements amount to significantly more than the abstract idea, (MPEP § 2106.05 sub I; see also MPEP § 2106.05(a) – (h)), and thus, the claim recites no more than the abstract idea. Thus, claim 14 is subject-matter ineligible.
Claim 16 depends directly or indirectly from claim 1. The claim recites more details or specifics to the additional element of “[(d)] . . . a second machine learning model,” where “[(d.2)] wherein the second machine learning model includes the plurality of encoding layers and a classifier head, [(d.2.1)] the classifier head configured to receive the layer output from a last encoding layer of the plurality of encoding layers of the second machine learning model and generate the prediction,” and accordingly, is merely more specific to the additional element. Thus, claim 16 is subject-matter ineligible.
Claim Rejections – 35 U.S.C. § 103
5. 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.
6. The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. § 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
7. This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. § 102(b)(2)(C) for any potential 35 U.S.C. § 102(a)(2) prior art against the later invention.
8. Claims 1, 2, 7-9, 14, and 16-20 are rejected under 35 U.S.C. § 103 as being unpatentable over US Published Application 20220277143 to Jayarao et al. [hereinafter Jayarao] in view of US Published Application 20220245349 to Pentyala et al. [hereinafter Pentyala].
Regarding claims 1 and 20, Jayarao teaches [a] method (Jayarao, Abstract, teaches “apparatus and methods for natural language understanding in conversational systems using machine learning processes”) of claim 1, and [a] non-transitory computer-readable storage medium storing a plurality of instructions executable by one or more processors, the plurality of instructions when executed by the one or more processors cause the one or more processors (Jayarao ¶ 0045 teaches “[i]nstruction memory 207 can store instructions that can be accessed (e.g., read) and executed by processors 201. For example, instruction memory 207 can be a non-transitory, computer-readable storage medium”) of claim 20, comprising:
[(a)] receiving data for use in generation of a machine learning model output (Jayarao ¶ 0020 teaches “machine learning processes may employ a natural language model, which is trained on retail data, and operates on input data characterizing textual information to determine the input data's intent [(that is, “input data” is receiving data for use in generation of a machine learning model output)]”;
[(b)] ingesting the data with a first machine learning model (Jayarao, Fig. 4, teaches “exemplary processing, by NLP computing device 102, of input textual data that may be received, for example, from a customer computing device 110, 112, 114 [Examiner annotations in dashed-line text boxes]”:
PNG
media_image1.png
380
783
media_image1.png
Greyscale
Jayarao ¶ 0061 teaches “NLP computing device 102 provides command 403 to each of the NLP model and the dependency embedding generation model. The NLP model may include a tokenizer 415, such as the WordPiece tokenizer, that generates tokens based on the command 403 [(that is, ingesting the data with a first machine learning model)]”) . . . ;
[(c)] receiving at least one intermediate output from the first machine learning model at a temporary store (Jayarao ¶ 0061 teaches “the NLP model [(that is, the first machine learning model)] includes a natural language model 419, such as a BERT model, that operates on the tokens generated by the tokenizer 415 to generate NLP output embeddings 421 [(that is, “embeddings” is receiving at least one intermediate output from the first machine learning model)]. In some examples, natural language model 419 generates output embeddings, and applies a linear layer to the output embeddings to generate NLP output embeddings 421”; Jayarao ¶ 0046 teaches “[p]rocessors 201 can also use working memory 202 to store dynamic data created during the operation of NLP computing device 102 [(that is, receiving . . . at a temporary store)]”),
[(c.1)] the at least one intermediate output comprising an output of an intermediate encoding layer in the first machine learning model (Jayarao, Fig. 6B, is a natural language processing architecture [Examiner annotation in dashed-line text boxes]:
PNG
media_image2.png
542
914
media_image2.png
Greyscale
Jayarao ¶ 0078 teaches the “[t]rained natural language model 604 may be, for example, a two layer or four layer DistilBERT model [(that is, an intermediate layer)]. The trained natural language model 604 generates a sequence output 605 for each token of the tokenized input data 601, and provides the sequence output 605 to linear layer 618 to generate natural language embeddings 621”;
[Examiner notes the plain meaning of the term “intermediate” is simply coming or occurring between two things. The broadest reasonable interpretation of the term “intermediate” covers a layer output “coming or occurring between two things” as a whole, which are the “command 403” and the “outputting 455” of the NLP computing device 102 of Jayarao, which is not inconsistent with the Applicant’s specification. (see, e.g., Specification ¶ 0025 & Fig. 1 (“each intermediate output is an output of one of the layers in the first model”))];
[Examiner also notes that the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a layer of a model, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to intermediate outputs of model layers]);
[(d)] ingesting the at least one intermediate output from the temporary store (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model to the concatenated embeddings [(that is, “apply . . . the concatenated embeddings” is ingesting the at least one intermediate output from the temporary store)]”) by another intermediate encoding layer in a second machine learning model (Jayarao, 6B, teaches “an architecture 650 for determining an intent 655 based on applying machine learning processes to input tokenized data 601 [Examiner annotations in dashed-line text boxes]:”
PNG
media_image3.png
907
1313
media_image3.png
Greyscale
Jayarao ¶ 0077 teaches “natural language embeddings 621 may then be concatenated with the dependency based embeddings 623 (e.g., at every time step) to generate concatenated embeddings 650. Output module 652 may receive concatenated embeddings 650, and apply a softmax function 653 [(that is, by another intermediate encoding layer in a second machine learning mode)]”), . . . ; and
[(e)] outputting a prediction with the second machine learning model (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model to the concatenated embeddings to generate intent and entity data 388, which identifies and characterizes a determined intent and entities [(that is, “intent and entity data” is outputting a prediction with the second machine learning model)]”;
[Examiner notes that the broadest reasonable interpretation of the terms “first machine learning model” and “second machine learning model” of the claims are defined by model components in relation to a GPU/CPU boundary, such as a GPU designating a first machine learning model and a CPU designating a second machine learning model, which is not inconsistent with the Applicant’s disclosure. (MPEP § 2111; see, e.g., Specification Fig. 3 and accompanying text)]).
Though Jayarao teaches a natural language understanding system that includes natural language processing (NLP) computing device employing first and second natural language model implemented with processors that can include one or more distinct processors, where the processors can include one or more central processing units (CPUs), one or more graphics processing units (GPUs); Jayarao, however, does not explicitly teach “the first machine learning model executed on a Graphic Processing Unit ("GPU")” and “the second machine learning model executed on the CPU.”
But Pentyala teaches a “NLP system 100 includes a first machine learning model 120 and one or more second machine learning model(s) 130A – N.” (Pentyala ¶ 0016). Generally, Pentyala teaches distinctions of a machine learning model being lighter (in terms of memory and processing resources needed for implementing the first machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while another second machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011);
[Examiner notes that the converse inherently holds, where a machine learning model being lighter (in terms of memory and processing resources needed for implementing the [second] machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while [a first] machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011)].
Jayarao and Pentyala are from the same or similar field of endeavor. Jayarao teaches a natural language processing system with a first model providing intermediate outputs to a second model. Pentyala teaches a natural language processing system implementing a layered or leveled system that includes a light machine learning model and a heavier machine learning model implemented using a CPU and a GPU, respectively.
Thus, it would have been obvious to a person having ordinary skill in the art as of the effective filing date of the Applicant’s invention to modify Jayarao pertaining to a NLP system with a first machine learning model providing intermediate outputs to a second machine learning model for a predictive output with the implementation of a first (heavyweight) model on a GPU and a second (lightweight) model on a CPU of Pentyala.
The motivation to do so is for a “mechanism [that] significantly speeds up the processing of utterances in an NLP when compared to existing NLP systems that rely on dedicated heavy machine learning models for processing the utterances.” (Pentyala ¶ 0013).
Regarding claim 2, the combination of Jayarao and Pentyala teaches all of the limitations of claim 1, as described above in detail.
Jayarao teaches -
wherein the first machine learning model comprises a first neural network (Jayarao ¶ 0056 teaches the “NLP model data 385 may further identify and characterize a corresponding linear layer, such as a linear neural network layer [(that is, the first machine learning model comprises a first neural network)]”), and wherein the second model comprises a second neural network (Jayarao ¶ 0039 teaches, relating to contextual intent detection, “a neural network is employed that takes into account a conversational history of a user, such as a customer or retailer associate, along with a latest intent, to proactively predict a next intent”).
Regarding claim 7, the combination of Jayarao and Pentyala teaches all of the limitations of claim 1, as described above in detail.
Jayarao teaches -
wherein the at least one intermediate output from the first machine learning model comprises a plurality of intermediate outputs,
each of the plurality of intermediate outputs being generated by a unique encoding layer of the first machine learning model (see above Jayarao, Fig. 6B; Jayarao ¶ 0061 teaches “the NLP model [(that is, the first machine learning model)] includes a natural language model 419, such as a BERT model, that operates on the tokens generated by the tokenizer 415 to generate NLP output embeddings 421 [(that is, “NLP output embeddings” is the at least one intermediate output from the first machine learning model comprises a plurality of intermediate outputs)]. In some examples, natural language model 419 generates output embeddings, and applies a linear layer [(that is, being generated by a unique encoding layer of the first machine learning model)] to the output embeddings to generate NLP output embeddings 421 [(that is, each of the plurality of intermediate outputs being generated by a unique encoding layer of the first machine learning model)]”; Jayarao ¶ 0075 & Fig. 6B teaches “[t]he trained natural language model 604 [(that is, the first machine learning model)] generates a sequence output 605 for each token of the tokenized input data 601, and provides the sequence output 605 to linear layer 618 to generate natural language embeddings 621”).
Regarding claim 8, the combination of Jayarao and Pentyala teaches all of the limitations of claim 7, as described above in detail.
Jayarao teaches -
wherein the second machine learning model comprises a plurality of encoding layers (Jayarao ¶ 0077 teaches “[o]utput module 652 may receive concatenated embeddings 650, and apply a softmax function 653 to the concatenated embeddings 650 to generate intent and entity tags 655 [(that is, the second machine learning model comprises a plurality of encoding layers )]”), each encoding layer of the plurality of encoding layers of the second machine learning model receiving an intermediate output from a corresponding intermediate encoding layer of the first machine learning model (Jayarao, Fig. 6B teaches
PNG
media_image4.png
972
1397
media_image4.png
Greyscale
Jayarao ¶ 0057 teaches “NLP computing device 102 may . . . apply the intent and entity classifier model [(that is, “653 and 655” are the second machine learning model)] to the concatenated embeddings [(that is, “concatenated embeddings” are receiving an intermediate output from a corresponding intermediate encoding layer of the first machine learning model)] to generate intent and entity data 388, which identifies and characterizes a determined intent and entities. NLP computing device 102 may store the intent and entity data 388 within database 116”);
[as discussed above, Examiner notes that the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a model layer, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to intermediate outputs of model layers]).
Regarding claim 9, the combination of Jayarao and Pentyala teaches all of the limitations of claim 8, as described above in detail.
Jayarao teaches -
further comprising:
generating, by each encoding layer of the plurality of encoding layers of the second machine learning model, a layer output based at least in part on the corresponding intermediate output of the intermediate encoding layer of the first machine learning model (Jayarao, 6B, teaches “an architecture 650 for determining an intent 655 based on applying machine learning processes to input tokenized data 601 [Examiner annotations in dashed-line text boxes]:”
PNG
media_image5.png
1002
1273
media_image5.png
Greyscale
Jayarao ¶ 0077 teaches “[o]utput module 652 [(that is, an intermediate encoding layer)] may receive concatenated embeddings 650, and apply a softmax function 653 [(that is, an intermediate encoding layer)] to the concatenated embeddings 650 [(that is, “applied” is generating, by each encoding layer of the plurality of encoding layers of the second machine learning model, a layer output of the identified next layer based at least in part on the corresponding intermediate output of the intermediate encoding layer of the first machine learning model)]”;
[see above, Examiner notes that the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a layer of a model, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to intermediate outputs of model layers]).
Regarding claim 14, the combination of Jayarao and Pentyala teaches all of the limitations of claim 9, as described above in detail.
Jayarao teaches –
further comprising:
generating, by an encoding layer of the plurality of encoding layers of the second machine learning model (Jayarao, Fig. 6B, teaches an architecture 650 having a second machine learning model receiving data from a first machine learning model to produce a “layer output” [Examiner annotations in dashed line text boxes]:”
PNG
media_image6.png
868
1310
media_image6.png
Greyscale
Jayarao ¶ 0077 teaches “output module 652 [(that is, second machine learning model )]” includes “a softmax function 653 [(that is, preceding encoding layer)]” and “generate intent and entity tags 655 [(that is, ”an encoding layer of the plurality of encoding layers of the second machine learning model)]”; see also above regarding Jayarao, Fig. 4 (“intent and entity classifier 388)), the layer output based at least in part on one or more outputs generated by one or more preceding encoding layers of the second machine learning model (Jayarao ¶ 0077 teaches “[o]utput module 652 may receive concatenated embeddings 650, and apply a softmax function 653 [(that is, preceding encoding layer)] to the concatenated embeddings 650 [from the first machine learning model] to generate intent and entity tags 655 [(that is, the layer output based at least in part on one or more outputs generated by one or more preceding encoding layers of the second machine learning model)]”;
[see above, Examiner notes that the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a layer of a model, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to intermediate outputs of model layers]).
Regarding claim 16, the combination of Jayarao and Pentyala teaches all of the limitations of claim 9, as described above in detail.
Jayarao teaches –
wherein the second machine learning model includes the plurality of encoding layers and a classifier head (Jayarao ¶ 0075 teaches “determining an intent 655 [(that is, “classifying,” which is a classifier head)] based on applying machine learning processes to input tokenized data 601 [from the natural language model 604]”; Jayarao ¶ 0077 teaches “apply softmax function 653 to . . . generate intent and entity tags 655 [(that is, the second machine learning model include the plurality of encoding layers and a classifier head)]), the classifier head configured to receive the layer output from a last encoding layer of the plurality of encoding layers of the second machine learning model (Jayarao ¶ 0077 teaches “[o]utput module 652 may receive concatenated embeddings 650, and apply a softmax function 653 [(that is, the last layer in the second model)] to the concatenated embeddings 650 to generate intent and entity tags 655 [(that is, “intent and entity tags” is generating the layer output o fa last layer in the second model)]”) and generate the prediction based on the layer output of the last layer (Jayarao ¶ 0077 teaches “apply softmax function 653 to . . . generate intent and entity tags 655 [(that is, generate the prediction based on the layer output of the last layer)]”).
Regarding claim 17, the combination of Jayarao and Pentyala teaches all of the limitations of claim 1, as described above in detail.
Pentyala teaches -
further comprising:
receiving the at least one intermediate output from the temporary store at a second Central Processing Unit (“second CPU”) (Jayarao ¶ 0043 teaches “[p]rocessors 201 can include one or more distinct processors, each having one or more cores. Each of the distinct processors can have the same or different structure. Processors 201 can include one or more central processing units (CPUs) [(that is, “one or more CPUs” is a first CPU and a second CPU)], one or more graphics processing units (GPUs), . . . and the like”; as described above in detail, Jayarao ¶ 0061 & Fig. 4, teaches “the NLP model [(that is, the first machine learning model)] includes a natural language model 419, such as a BERT model, that operates on the tokens generated by the tokenizer 415 to generate NLP output embeddings 421 [(that is, “embeddings” is receiving at least one intermediate output from the temporary store at a second Central Processing Unit (“second CPU))]”);
ingesting the at least one intermediate output with a . . . machine learning model (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model [(that is, a . . . machine learning model )] to the concatenated embeddings [(that is, ingesting the at least one intermediate output)]”) . . . ;
* * *
Though Jayarao teaches the applying a machine learning model to an embedding, or intermediate output, Jayarao, however, does not explicitly teach a “third machine learning model on the second CPU.”
But Pentyala teaches –
* * *
[ingesting] . . . with a third machine learning model on the second CPU (Pentyala, Fig. 1, teaches second machine learning model(s) 130A-N [Examiner annotations in dashed-line text boxes]:
PNG
media_image7.png
458
835
media_image7.png
Greyscale
Pentyala ¶ 0017 teaches “[t]he [NLP] system 100 may include one or multiple ones of the second machine learning models 130A-N. A second machine learning model can be of a predetermined type, i.e., dedicated to a particular natural language task. In some implementations, two second machine learning models can be of different types (i.e., dedicated to different NLP tasks). Additionally or alternatively, two second machine learning models can be applicable to different fields”; Pentyala ¶ 0027 teaches “the dedicated natural language task is one of intent detection (e.g., 130A) [(that is, “model 130A” is a second machine learning model)], named entity recognition (NER 130D) [(that is, “model 130D” is a third machine learning model)] . . . .”;
Pentyala also teaches a “NLP system 100 includes a first machine learning model 120 and one or more second machine learning model(s) 130A – N.” (Pentyala ¶ 0016). Generally, Pentyala teaches distinctions of a machine learning model being lighter (in terms of memory and processing resources needed for implementing the first machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while another second machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011);
[Examiner notes that the converse inherently holds, where a machine learning model being lighter (in terms of memory and processing resources needed for implementing the [second] machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while [a first] machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011) [(that is, a third machine learning model on a second CPU)]); and
outputting a prediction with the third machine learning model (Pentyala ¶ 0027 teaches an “utterance is inputted to the second machine learning model (e.g., one or multiple ones of the second machine learning models 130A-N) to obtain an output [455] of the dedicated natural language task [(that is, outputting a prediction with the third machine learning model)]”).
Jayarao and Pentyala are from the same or similar field of endeavor. Jayarao teaches a natural language processing system with a first model providing intermediate outputs to a second model. Pentyala teaches a natural language processing system implementing a layered or leveled system that includes a light machine learning model and a heavier machine learning model implemented using a CPU and a GPU, respectively. Thus, it would have been obvious to a person having ordinary skill in the art as of the effective filing date of the Applicant’s invention to modify Jayarao pertaining to a NLP system with a first machine learning model providing intermediate outputs with a third machine learning model for a predictive output, where the third model is a lightweight model on a second CPU of Pentyala.
The motivation to do so is for a “mechanism [that] significantly speeds up the processing of utterances in an NLP when compared to existing NLP systems that rely on dedicated heavy machine learning models for processing the utterances.” (Pentyala ¶ 0013).
Regarding claim 18, the combination of Jayarao and Pentyala teaches all of the limitations of claim 17, as described above in detail.
wherein the third machine learning model comprises only a classifier head (Pentyala ¶ 0027 teaches “the dedicated natural language task is one of intent detection (e.g., 130A) [(that is, “model 130A” is a second machine learning model)], named entity recognition (NER 130D) [(that is, “NER 130D” is only a classifier head )] . . . .”).
Regarding claim 19, Jayarao teaches [a] system (Jayarao, abstract, teaches “apparatus and methods for natural language understanding in conversational systems using machine learning processes”) comprising:
[(a)] memory comprising a temporary store (Jayarao ¶ 0046 teaches “[p]rocessors 201 can also use working memory 202 [(that is, memory comprising a temporary store)] to store dynamic data created during the operation of NLP computing device 102”);
[(b) a Graphics Processing Unit machine running] a first machine learning model . . . (Jayarao, Fig. 4, teaches “exemplary processing, by NLP computing device 102 [Examiner annotations in dashed-line text boxes]:”
PNG
media_image8.png
430
791
media_image8.png
Greyscale
Jayarao ¶ 0060 “input textual data that may be received for example, from a customer computing device 110, 112, 114”) configured to:
[(b.1)] receive data for use in generation of a machine learning model output (Jayarao ¶ 0020 teaches “machine learning processes may employ a natural language model, which is trained on retail data, and operates on input data characterizing textual information to determine the input data's intent [(that is, “input data” is receive data for use in generation of a machine learning model output)]”);
[(b.2)] ingest the data with the first machine learning model (Jayarao ¶ 0061 teaches “NLP computing device 102 provides command 403 to each of the NLP model and the dependency embedding generation model. The NLP model may include a tokenizer 415, such as the WordPiece tokenizer, that generates tokens based on the command 403 [(that is, ingest the data with the first machine learning model)]”);
[(b.3)] generate at least one intermediate output from the first machine learning model (Jayarao ¶ 0061 teaches “the NLP model [(that is, the first machine learning model)] includes a natural language model 419, such as a BERT model, that operates on the tokens generated by the tokenizer 415 to generate NLP output embeddings 421 [(that is, “embeddings” is generate at least one intermediate output from the first machine learning model)), the at least one intermediate output comprising an output of an intermediate encoding layer in the first machine learning model (Jayarao, Fig. 6B, is a natural language processing architecture [Examiner annotation in dashed-line text boxes]:
PNG
media_image3.png
907
1313
media_image3.png
Greyscale
Jayarao ¶ 0078 teaches the “[t]rained natural language model 604 may be, for example, a two layer or four layer DistilBERT model. The trained natural language model 604 generates a sequence output 605 for each token of the tokenized input data 601, and provides the sequence output 605 to linear layer 618 to generate natural language embeddings 621 [(that is, generate at least one intermediate output from the first machine learning model)]”;
[Examiner notes the plain meaning of the term “intermediate” is coming or occurring between two things. The broadest reasonable interpretation of the term “intermediate” covers a layer output “coming or occurring between two things” as a whole, which are the “command 403” and the “outputting 455” of the NLP computing device 102 of Jayarao, which is not inconsistent with the Applicant’s specification. (see, e.g., Specification ¶ 0025 & Fig. 1 (“each intermediate output is an output of one of the layers in the first model”))];
[Examiner also notes that the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a layer of a model, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to model layer outputs]); and
[(b.4)] provide the at least one intermediate output to the temporary store (Jayarao ¶ 0061 teaches “natural language model 419 generates output embeddings, and applies a linear layer to the output embeddings to generate NLP output embeddings 421”; Jayarao ¶ 0046 teaches “[p]rocessors 201 can also use working memory 202 to store dynamic data created during the operation of NLP computing device 102 [(that is, provide the at least one intermediate output to the temporary store)]”);
[(c) a Central Processing Unit machine running] a second machine learning model . . . configured to (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model [(that is, a second machine learning model )] to the concatenated embeddings”):
[(c.1)] receive the at least one intermediate output from the temporary store (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model to the concatenated embeddings [(that is, “apply . . . the concatenated embeddings” is receive the at least one intermediate output from the temporary store)]”);
[(c.2)] ingest the at least one intermediate output from the temporary store by another intermediate encoding layer in the second machine learning model (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model [(that is, a second machine learning model )] to the concatenated embeddings”; Jayarao ¶ 0074 teaches that “[o]utput module 652 may receive concatenated embeddings 650 [from working memory 202 (that is, the temporary store)], and generate output 653 [(that is, by another intermediate encoding layer)], which includes a next sentence prediction value (NSP), and masked language modeling prediction values (MLP) [(that is, ingest the at least one intermediate output from the temporary store by another intermediate encoding layer in the second machine learning model)]”); and
[(c.3)] output a prediction with the second machine learning model (Jayarao ¶ 0057 teaches “apply the intent and entity classifier model to the concatenated embeddings to generate intent and entity data 388, which identifies and characterizes a determined intent and entities [(that is, “intent and entity data” is output a prediction with the second machine learning model)]”).
Though Jayarao teaches a natural language understanding system that includes natural language processing (NLP) computing device employing first and second natural language model implemented with processors that can include one or more distinct processors, where the processors can include one or more central processing units (CPUs), one or more graphics processing units (GPUs); Jayarao, however, does not explicitly teach “a first machine learning model on a Graphic Processing Unit ("GPU")” and “a second machine learning model] on the CPU.”
But Pentyala teaches a “NLP system 100 includes a first machine learning model 120 and one or more second machine learning model(s) 130A – N.” (Pentyala ¶ 0016). Generally, Pentyala teaches distinctions of a machine learning model being lighter (in terms of memory and processing resources needed for implementing the first machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while another second machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011);
[Examiner notes that the converse inherently holds, where a machine learning model being lighter (in terms of memory and processing resources needed for implementing the [second] machine learning model), more efficient (faster) supporting CPU only processing (as opposed to needing support of dedicated processing units such as graphics processing units (GPUs) in addition to CPU processing), while [a first] machine learning model is more specialized (e.g., defined for a particular task, or a particular field (medical, financial, etc.)), uses more compute and storage resources, and is slower. In some implementations, the heavyweight machine learning model can be implemented on GPUs and/or a combination of CPU/GPUs. (see Pentyala ¶ 0011);.
Jayarao and Pentyala are from the same or similar field of endeavor. Jayarao teaches a natural language processing system with a first model providing intermediate outputs to a second model. Pentyala teaches a natural language processing system implementing a layered or leveled system that includes a light machine learning model and a heavier machine learning model implemented using a CPU and a GPU, respectively. Thus, it would have been obvious to a person having ordinary skill in the art as of the effective filing date of the Applicant’s invention to modify Jayarao pertaining to a NLP system with a first machine learning model providing intermediate outputs to a second machine learning model for a predictive output with the implementation of a first (heavyweight) model on a GPU and a second (lightweight) model on a CPU of Pentyala.
The motivation to do so is for a “mechanism [that] significantly speeds up the processing of utterances in an NLP when compared to existing NLP systems that rely on dedicated heavy machine learning models for processing the utterances.” (Pentyala ¶ 0013).
9. Claims 3-6 are rejected under 35 U.S.C. § 103 as being unpatentable over US Published Application 20220277143 to Jayarao et al. [hereinafter Jayarao] in view of US Published Application 20220245349 to Pentyala et al. [hereinafter Pentyala], and US Published Application 20210383272 to Hua et al. [hereinafter Hua].
Regarding claim 3, the combination of Jayarao and Pentyala teaches all of the limitations of claim 2, as described above in detail.
Though Jayarao and Pentyala teach models that are implemented using neural networks, the combination of Jayarao and Pentyala does not explicitly teach –
wherein the first neural network comprises a deep learning neural network.
But Hua teaches -
wherein the first neural network comprises a deep learning neural network (Hua ¶ 0003 teaches “Deep Neural Networks (DNNs) have become the most successful machine-learning technique for solving a variety of tasks, including classification, natural language understanding (NLU), etc. [(that is, the first neural network comprises a deep learning neural network)]”).
Jayarao, Pentyala, and Hua are from the same or similar field of endeavor. Jayarao teaches a natural language processing system with a first model providing intermediate outputs to a second model. Pentyala teaches a natural language processing system implementing a layered or leveled system that includes a light machine learning model and a heavier machine learning model implemented using a CPU and a GPU, respectively. Hua teaches [a] deep learning neural network learns to map a set of inputs to a set of outputs from training data (including input variables and target variables) via an optimization process that requires one or more loss functions to calculate the model error(s). Thus, it would have been obvious to a person having ordinary skill in the art as of the effective filing date of the Applicant’s invention to modify the combination of Jayarao and Pentyala pertaining to a NLP system with a first machine learning model providing intermediate outputs to a second machine learning model for a predictive output, the neural networks providing an implementation of a first (heavyweight) model on a GPU and a second (lightweight) model on a CPU with the deep neural network of Hua.
The motivation to do so is because “Deep Neural Networks (DNNs) have become the most successful machine-learning technique for solving a variety of tasks, including classification, natural language understanding (NLU), etc.” (Hua ¶ 0003).
Regarding claim 4, the combination of Jayarao, Pentyala, and Hua teaches all of the limitations of claim 3, as described above in detail.
Hua teaches -
wherein the deep learning neural network comprises a transformer (Hua ¶ 0091 teaches that “[f]or DNN based classification models including a feature extractor and classifier, Fisher information of parameters of both the feature extractor and classifier may be calculated. In one embodiment utilizing Bidirectional Encoder Representations from Transformers (BERT) [(that is, a transformer)] for text classification [(that is, wherein the deep learning neural network comprises a transformer)]”).
Regarding claim 5, the combination of Jayarao, Pentyala, and Hua teaches all of the limitations of claim 4, as described above in detail.
Hua teaches -
wherein the transformer comprises a Bidirectional Encoder Representations from Transformers (“BERT”) model (Hua ¶ 0091 teaches that “[f]or DNN based classification models including a feature extractor and classifier, Fisher information of parameters of both the feature extractor and classifier may be calculated. In one embodiment utilizing Bidirectional Encoder Representations from Transformers (BERT) for text classification [(that is, wherein the transformer comprises a Bidirectional Encoder Representations from Transformers (“BERT”) model)]”).
Regarding claim 6, the combination of Jayarao, Pentyala, and Hua teaches all of the limitations of claim 3, as described above in detail.
Jayarao teaches -
wherein the second machine learning model comprises a task specific model (Jayarao ¶¶ 0057 teaches “NLP computing device 102 may concatenate the output embeddings from each of the NLP model and the dependency embedding generation model , and apply the intent and entity classifier model to the concatenated embeddings to generate intent and entity data 388 , which identifies and characterizes a determined intent and entities [(that is, “identifies and characterizes” is wherein the second machine learning model comprises a task specific model)]”).
Response to Arguments
10. Examiner has fully considered the Applicant’s arguments, and responds below accordingly.
Section 101
11. Applicant submits that “Claims 9-16 are rejected under 35 U.S.C. § 101 for allegedly being directed to an abstract idea without significantly more. Applicant respectfully submits that the present amendment to claims renders this rejection moot. For at least these reasons, the Applicant respectfully requests withdrawal of the§ 101 rejection of the pending claims.” (Response at p. 7).
Examiner’s Response:
Examiner respectfully disagrees because a dependent claim can also introduce a limitation directed to an abstract idea and be rejected under 35 USC 101, even though the independent claim is considered eligible under 35 USC 101.
Under Rule 1.75, “[o]ne or more claims may be presented in dependent form, referring back to and further limiting another claim or claims in the same application . . . . Claims in dependent form shall be construed to include all the limitations of the claim incorporated by reference into the dependent claim. A multiple dependent claim shall be construed to incorporate by reference all the limitations of each of the particular claims in relation to which it is being considered.” (37 CFR 1.75; MPEP § 608.01(n)).
Accordingly, as set out above in detail, claims 9, 14, and 16 depend directly or indirectly from claim 1 introduce a limitation directed to an abstract idea under Step 2A Prong One. Further, under Step 2A Prong Two, the additional elements (e.g., graphic processing unit, temporary store, central processing unit) are generic computer components used to implement the abstract idea, and additional elements of “receiving,” “ingesting,” and “outputting” are insignificant extra-solution activities of mere data gathering, (MPEP § 2106.05(g)), which does not integrate the abstract idea into a practical application, as set out above in detail. Under Step 2B, the additional elements (e.g., graphic processing unit, temporary store, central processing unit) are generic computer components used to implement the abstract idea and additional elements of “receiving,” “ingesting,” and “outputting” are well-understood, routine, and conventional activities that do not amount to significantly more than the abstract idea, as set out above in detail.
Therefore, for these reasons, claims 9, 14, and 16 are subject-matter ineligible.
Section 103
12. Applicant submits that “Jayarao teaches the use of outputs of the machine learning models, referring to these outputs as embeddings. To wit, "the output of each of the natural language model and the dependency embedding generation model." Jayarao, Paragraph [0037]. Jayarao teaches that these outputs are generated by passing through the entire model. See, Jayarao, Paragraph [0039]. Jayarao does not anywhere teach that these embeddings are "an output of an intermediate layer in the first machine learning modef' as recited in claim 1.” (Response at p. 8).
Also, “to expedite prosecution Applicant have amended the independent claims to recite
* * *
[(d)] ingesting the at least one intermediate output from the temporary store by another intermediate encoding layer in a second machine learning model, wherein the second machine learning model is executed on a Central Processing Unit;
* *
[(claim 1, lines 8-10)]. Specifically, an intermediate encoding layer of the second machine learning model receives (as input), the output generated by a corresponding intermediate encoding layer of the first machine learning model. Applicant respectfully submit that Jayarao simply does not describe or suggest, at least, this amended feature.” (Response at p. 8).
Examiner’s Response:
Examiner respectfully disagrees because the claim is not so limited as suggested by Applicant.
As explained above in detail, the plain meaning of the term “intermediate” is coming or occurring between two things. The broadest reasonable interpretation of the term “intermediate” covers a layer output “coming or occurring between two things” as a whole, which are the “command 403” and the “outputting 455” of the NLP computing device 102 of Jayarao, which is not inconsistent with the Applicant’s specification. (see, e.g., Specification ¶ 0025 & Fig. 1 (“each intermediate output is an output of one of the layers in the first model”)).
Also, the broadest reasonable interpretation of the claim term “intermediate encoding layer” or “encoding layer” is simply that of a model layer, which is not inconsistent with the Applicant’s disclosure, (MPEP § 2111; see Specification ¶ 0035 (“each intermediate output is an output of one of the layers in the first model. In some embodiments, some or all of the layers in the first model each generates an intermediate output”)), and accordingly, covers the teachings of Jayarao relating to intermediate outputs of model layers.
Also, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. Where a rejection of a claim is based on two or more references, a reply that is limited to what a subset of the applied references teaches or fails to teach, or that fails to address the combined teaching of the applied references may be considered to be an argument that attacks the reference(s) individually, as is the case here with the cited prior art of Jayarao. MPEP § 2145.IV.
Accordingly, the prior art of Jayarao teaches the features of Applicant’s claims, as described above in detail.
Conclusion
13. The prior art made of record and not relied upon is considered pertinent to Applicant's disclosure:
(Pati et al., "Demystifying BERT: Implications for Accelerator Design," arXiv (April 2021)) focuses on BERT, one of the most popular NLP transfer learning algorithms, to identify how BERT’s algorithmic behavior can guide future accelerator design. Many accelerators, including GPUs and TPUs are being used to train BERT. In this study we choose to study BERT on a GPU because of its wide availability as well as its popularity for DNN training. However, our key takeaways are accelerator agnostic and should be applicable to other accelerators (discussed further in Section 5). We study BERT pre-training on a system consisting of an AMD Ryzen™ Threadripper™ CPU [7] and an AMD Instinct™ MI100 GPU [10] with 32GB HBM2 [39].
(US Published Application 20210374947 to Shin et al.) teaches to facilitate generation of one medical image from another medical image using one or more neural networks trained using a generative adversarial network (GAN) that utilizes a bidirectional encoder representations from transformers (BERT) as a discriminator. In at least one embodiment, one or more neural networks trained using a GAN comprising a BERT discriminator generate a positron emission tomography (PET) image from a magnetic resonance imaging (MRI) image.
14. Any inquiry concerning this communication or earlier communications from the Examiner should be directed to KEVIN L. SMITH whose telephone number is (571) 272-5964. Normally, the Examiner is available on Monday-Thursday 0730-1730.
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, KAKALI CHAKI can be reached on 571-272-3719. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/K.L.S./
Examiner, Art Unit 2122
/KAKALI CHAKI/Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights