DETAILED ACTION
Notice of 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 .
Response to Amendment
Applicant’s amendment and remarks dated 9/8/2026 have been considered. Claims 3-4, 6-7, and 15-16 are cancelled. Claims 1-2, 5, 8-14, and 17-20 are pending.
Drawing Objections. The objection to Fig. 5 is withdrawn in view of the substitute drawing provided by Applicant.
35 U.S.C. 101 Rejections. The rejections to claims 17-20 under 35 U.S.C. 101 for not pertaining to an eligible subject matter category are withdrawn in view of Applicant’s amendments to claim 17. However, all rejections under 35 U.S.C. 101 for being directed to an abstract idea without significantly more are being maintained as set forth in the detailed rejections below.
Response to Arguments
On page 7 of Applicant’s 9/8/2026 Amendment and remarks, Applicant asserts that at least paras. 0072, 00105-06, 00108, 00155, and 00119-20 provide sufficient written description support for the claim amendments.
The examiner agrees that such portions of the specification, together with at least original claims 3-4 and 6-7, provide sufficient written description support for the claim amendments.
On page 8 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 101, with respect to Step 2A, Prong 1, Applicant argues:
PNG
media_image1.png
336
638
media_image1.png
Greyscale
The examiner respectfully disagrees. While the “each of the plurality of AI operators is preconfigured on the database system” is not a mental step (and is therefore analyzed under Step 2A, Prong 2, and Step 2B), the remaining limitation remains a mental step, where a human can consider the plurality of AI operators and their estimated execution cost when determining an execution plan.
In response to Applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., access to the database systems’ internal cost model for each preconfigured AI operator”) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
On pages 8-9 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 101, with respect to Step 2A, Prong 1, Applicant argues:
PNG
media_image2.png
230
660
media_image2.png
Greyscale
PNG
media_image3.png
52
618
media_image3.png
Greyscale
The examiner respectfully disagrees. In response to Applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., real-time access to the database system’s resource availability and cost”) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993). A human can mentally determine how to apportion the N training sample groups into M sample groups based on the information recited. The claim does not require access to any “cost models” and simply requires cost estimations, which can be based on a human’s education and experience.
On page 9 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claims 1 and 17 under 35 U.S.C. 101, with respect to Step 2A, Prong 1, Applicant argues that the “separately performing parallel training on the to-be-trained model…” limitation is not a mental step.
The examiner agrees that performing parallel training is not a mental step, and this limitation (as actually recited and not paraphrased by Applicant) has been addressed under Step 2A, Prong 2 and Step 2B.
On pages 9-11 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejections of the independent claims under 35 U.S.C. 101, with respect to Step 2A, Prong 2, Applicant argues that para. 0004 identifies a problem with conventional approaches, and that the claims address the technical problem by performing certain limitations.
The examiner respectfully disagrees. MPEP 2106.05(a) explains that “the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements.” As explained above, the “generating an execution plan” and the “obtaining, … M training sample groups…” limitations are mental processes that cannot provide the improvement alone. Moreover, the “training a to-be-trained model” limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Claim 17, which is nearly identical to original claim 17, remains rejected for the same reasons explained herein. Similarly, the steps of “train a to-be-trained model” and “update the to-be-trained mode” are recited at a high-level of generality and amount to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitations merely provide instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. (See MPEP 2106.05(f)).
On page 11, Applicant argues that the claims are analogous to Example 47, claim 3. The examiner respectfully disagrees. Example 47, claim 3 provides an improvement in the technical field of network intrusion detection. In contrast, as explained above, the present claims do not provide any technical improvement to any technology or technical field. The use of generic computing components and generic ML models merely improve the mental processes identified by the office action.
On page 11, the examiner respectfully disagrees with Applicant’s characterization that the “core of claim 1” is “the database systems’ generation of an execution plan.” The claims do not require the database system to generate an execution plan, and do not even require the database system to execute such plan. The claims merely require the execution plan to be “generated”, and the scope of the claims includes generation in the human mind.
On pages 11-12 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejections of the independent claims under 35 U.S.C. 101, with respect to Step 2B, Applicant generally objects to the office actions’s reliance on the “well-understood, routine, and conventional activity” consideration with respect to the data gathering steps in the claims.
The examiner respectfully disagrees. As set forth by MPEP 2144.04, “an examiner may utilize legal precedent as a source of supporting rationale when warranted and appropriately supported.” Here, the examiner has identified several legal precedents set forth in MPEP 2106.05(d) that generally represent that data gathering steps are well-understood, routine, and conventional activity.
Moreover, the examiner’s characterization of certain steps as mere “data gathering” are supported by numerous examples in the USPTO’s Eligibility Examples, including at least Examples 47-48. For example, at page 21 of the July 2024 SME Updates with respect to Examples 47-49, explains that data gathering is well-understood, and similarly, para. 0003 of the instant specification sets forth in the Background that obtaining knowledge using computers is known.
On pages 13-14 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 103, Applicant argues that the TONG reference’s disclosure of a “maximum cost constraint” does not disclose the “estimated execution cost” limitation. Applicant argues that the “maximum cost” is a “user-specified configuration constraint” and that somehow this distinguishes from the claimed “estimated execution cost” that is “obtained based on estimated execution costs of the plurality of AI operators.”
The examiner respectfully disagrees. As amended, the “obtained based on the estimated execution costs of the plurality of AI operators” simply moves up subject matter from now-cancelled claim 4, and in the previous office action, this limitation was taught by the combination of the TONG, KUSHNIR, and ALONSO references, where now one of ordinary skill could determine a maximum cost based on the estimated execution cost of particular algorithms, such as SGD, ADAM, XGBoost, as disclosed by ALONSO. Therefore, in response to Applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Moreover, nothing in the claims precludes the “estimated execution cost” from being obtained from a user, where the user considers the individual execution costs of individual operators.
Further, nothing in the claims precludes the “estimated execution cost” from being a conservative, “max cost constraint”, where the individual execution costs are considered as part of a maximum cost constraint. The specification and claims do not provide a specific definition for “estimated execution cost”, and after the amendment, the only requirement is that such estimated execution cost is “based on estimated execution costs of the plurality of AI operators”, and therefore a maximum constraint that takes such costs into account would read on the “estimated execution costs” limitation.
On page 14 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 103, Applicant argues with respect to the KUSHNIR reference:
PNG
media_image4.png
496
658
media_image4.png
Greyscale
The examiner respectfully disagrees. KUSHNIR stands for the proposition of selecting a number of data points based on a training budget, where in KUSHNIR an expensive operation was data labeling. The combination of TONG and KUSHNIR now utilizes the overall budget teachings of KUSHNIR to determine a subset of data points for training as explained in the office action. Applicant’s argument relies on its assertion that TONG does not teach the “estimated execution cost” and that a maximum cost, or a budget, cannot read on the “estimated execution cost” limitation, which as explained above, the examiner respectfully disagrees with.
On pages 14-15 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 103, Applicant argues that the prior art reference does not expressly teach the “wherein a value of M is negatively correlated” and a “value of M is positively correlated limitations.”
The examiner respectfully disagrees. MPEP 2144.01 explains that in “considering the disclosure of a reference, it is proper to take into account not only specific teachings of the reference but also the inferences which one skilled in the art would reasonably be expected to draw therefrom. Here, the value of “M” sized based on the budget teachings of KUSHNIR, and one of ordinary skill would understand that if such budget relates to training costs (as in TONG), then as the training costs increases, the number of samples used (M) would necessarily need to go down. In contrast, if the number of computation resources goes up, so can the number of samples used for training. These are natural inferences based on the type of training budget.
Moreover, the claims themselves merely require M to be negatively/positively correlated with certain data points. The claims do not require actively selecting the value of M in order to optimize any particular criteria.
On page 15 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 1 under 35 U.S.C. 103, Applicant argues with respect to the ALONSO reference:
PNG
media_image5.png
404
662
media_image5.png
Greyscale
In response to Applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986). As explained on pages 32-33 and 36-37 of the 6/8/2026 office action, this limitation is taught by the combination of the TONG, KUSHNIR, and ALONSO references, where ALONSO teaches the concept of individual AI operators (e.g., SGD, ADAM, and XGBoost), and TONG teaches the concept of a maximum cost, where one of ordinary skill would modify the execution plan of TONG to consider execution costs of the individual AI operators when determining the maximum cost of TONG.
On page 16 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of claim 17 under 35 U.S.C. 103, Applicant argues that the KUSHNIR reference does not teach the “M training sample groups” or “M pieces of parameter update data” limitations.
Applicant’s arguments are not persuasive for the same reasons explained above with respect to claim 1. KUSHNIR teaches the general concept of selecting a number of training groups (M), and a natural inference is that if there are M inputs to a machine learning model, than there are also M outputs (“obtain M pieces of parameter update data”) based on the M number of inputs.
On pages 15-17 of Applicant’s 9/8/2026 Amendment and remarks, with respect to the rejection of the remaining dependent claims under 35 U.S.C. 103, Applicant argues that such claims should be allowed for the same reasons argued with respect to the independent claims.
The examiner respectfully disagrees for the same reasons explained above with respect to the independent claims.
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 1-2, 5, 8-14, and 17-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding Step 1 of the Alice/Mayo framework, Claims 1-2, 5, and 8-12 are directed to a method (a process), and claims 13-14 are directed to an apparatus (a machine), and claims 17-209 are directed to a non-transitory computer readable medium (an article of manufacture), which each fall within one of the four statutory categories of inventions.
Regarding Claim 1
Step 2A, prong 1 (Is the claim directed to a law of nature, a natural phenomenon or an abstract idea).
Claim 1 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including an observation, evaluation, judgment, opinion) or with the aid of pencil and paper but for the recitation of generic computer components (e.g., “database system”, “to-be-trained model”, and “trained model).
generating an execution plan of the model training policy and an estimated execution cost of the execution plan, wherein the execution plan comprises a plurality of AI operators … and the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators; (under the broadest reasonable interpretation, a human can mentally determine a plan for training a machine learning model, such as a plan to train a model for a particular task, using backpropagation with SGD, and a human can mentally estimate the resource cost of such plan (e.g., needing 12 hours to train on available servers), where such plan includes estimates based on the number of AI operators and their expected execution costs)
obtaining, based on the estimated execution cost, and a currently available computing resource of the database system, M training sample groups in the N training sample groups, wherein a value of M is negatively correlated with the estimated execution cost and positively correlated with the currently available computing resource; (under the broadest reasonable interpretation, a human can mentally determine a subset of the training samples to use for training based on the estimated execution cost that the human wants to meet, where a human can mentally determine M, as a subset of the entire training set, where M goes down as the estimated execution cost goes up, because more training samples requires more training resources, and where M goes up as the available resources for training goes up)
Step 2A, prong 2 (Does the claim recite additional elements that integrate the judicial exception into a practical application?).
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional elements (e.g., “database system”, “to-be-trained model”, and “trained model) which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “data processing method, applied to a database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of a database system. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (a database system). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “obtaining a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups” limitation, such additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Regarding the “executed by the database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of a database system. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (a database system). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “an operator type of each of the plurality of AI operators is preconfigured on the database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of a database system. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (a database system). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “training a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of generic training of a model. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (generic training of a model). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of generic training of a model. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (generic training of a model). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, after considering all claim elements individually and as an ordered combination, it is determined that the claims do not integrate the judicial exception into a practical application.
Step 2B (Does the claim recite additional elements that amount to significantly more than the judicial exception?)
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional elements (e.g., “database system”, “to-be-trained model”, and “trained model) are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “data processing method, applied to a database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “obtaining a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups” limitation, as discussed above, the additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Regarding the “executed by the database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “an operator type of each of the plurality of AI operators is preconfigured on the database system” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “training a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Accordingly, at Step 2B after considering all claim elements individually and as an ordered combination, it is determined that the claims do not integrate the judicial exception into a practical application.
Regarding Claim 2
Step 2A, Prong 2
Regarding the “wherein the parameter update data is an update gradient of the to-be-trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of generic training techniques. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (generic training techniques). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Moreover, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use (training using gradients, such as SGD). As explained by the Supreme Court, a claim directed to a judicial exception cannot be made eligible "simply by having the applicant acquiesce to limiting the reach of the patent for the formula to a particular technological use." Diamond v. Diehr, 450 U.S. 175, 192 n.14, 209 USPQ 1, 10 n. 14 (1981). Thus, limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not integrate a judicial exception into a practical application.
Step 2B
Regarding the “wherein the parameter update data is an update gradient of the to-be-trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Moreover, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use as explained above, which does not amount to significantly more than the judicial exception. MPEP 2106.05(h).
Regarding Claim 5
Step 2A, Prong 1
obtaining the estimated execution costs of the plurality of AI operators (under the broadest reasonable interpretation, a human can estimate the execution cost based on the AI operators, such as estimating the amount of resources and time to train a model with a particular number of samples, using SGD and ADAM, based on personal experience)
Step 2A, Prong 2
Regarding the “by using an execution plan query statement” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of a basic SQL database command. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (a basic SQL database command). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Step 2B
Regarding the “by using an execution plan query statement” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding Claim 8
Step 2A, Prong 2
Regarding the “performing a shuffle operation on the plurality of training samples, to obtain a plurality of shuffled training samples; and grouping the plurality of shuffled training samples, to obtain the N training sample groups” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation attempts to cover a solution to an identified problem with no restriction on how the result is accomplished, or provides no description of the mechanism for accomplishing the result. Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Step 2B
Regarding the “performing a shuffle operation on the plurality of training samples, to obtain a plurality of shuffled training samples; and grouping the plurality of shuffled training samples, to obtain the N training sample groups” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation attempts to cover a solution to an identified problem with no restriction on how the result is accomplished, or provides no description of the mechanism for accomplishing the result. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding Claim 9
Step 2A, Prong 1
obtaining ... an execution cost of each of the X trained models (under the broadest reasonable interpretation, a human can mentally estimate the execution cost for a particular model based on personal experience)
determining a target model from the X trained models (under the broadest reasonable interpretation, a human can mentally determine a particular model from a set of models)
Step 2A, Prong 2
Regarding the “wherein there are X to-be-trained models in a one-to-one correspondence to X trained models, the X trained models are different models used to implement a target task” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of trained models. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (trained models). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “after the updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model, the method further comprises: obtaining a model inference request indicating the target task; obtaining the currently available computing resource” limitation, such additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Regarding the “performing model inference by using the target model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of model inference. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (model inference). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Step 2B
Regarding the “wherein there are X to-be-trained models in a one-to-one correspondence to X trained models, the X trained models are different models used to implement a target task” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “after the updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model, the method further comprises: obtaining a model inference request indicating the target task; obtaining the currently available computing resource” limitation, as discussed above, the additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Regarding the “performing model inference by using the target model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding Claim 10
Step 2A, Prong 2
Regarding the “wherein the computing resource comprises: an input/output (I/O) resource, a central processing unit (CPU) resource, a graphics processing unit (GPU) resource, and/or a memory resource” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use. As explained by the Supreme Court, a claim directed to a judicial exception cannot be made eligible "simply by having the applicant acquiesce to limiting the reach of the patent for the formula to a particular technological use." Diamond v. Diehr, 450 U.S. 175, 192 n.14, 209 USPQ 1, 10 n. 14 (1981). Thus, limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not integrate a judicial exception into a practical application.
Step 2B
Regarding the “wherein the computing resource comprises: an input/output (I/O) resource, a central processing unit (CPU) resource, a graphics processing unit (GPU) resource, and/or a memory resource” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use as explained above, which does not amount to significantly more than the judicial exception. MPEP 2106.05(h).
Regarding Claim 11
Step 2A, Prong 2
Regarding the “wherein the model inference request is a structured query language (SQL) statement” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use. As explained by the Supreme Court, a claim directed to a judicial exception cannot be made eligible "simply by having the applicant acquiesce to limiting the reach of the patent for the formula to a particular technological use." Diamond v. Diehr, 450 U.S. 175, 192 n.14, 209 USPQ 1, 10 n. 14 (1981). Thus, limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not integrate a judicial exception into a practical application.
Step 2B
Regarding the “wherein the model inference request is a structured query language (SQL) statement” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use as explained above, which does not amount to significantly more than the judicial exception. MPEP 2106.05(h).
Regarding Claim 12
Step 2A, Prong 2
Regarding the “wherein the model training request is an SQL statement” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use. As explained by the Supreme Court, a claim directed to a judicial exception cannot be made eligible "simply by having the applicant acquiesce to limiting the reach of the patent for the formula to a particular technological use." Diamond v. Diehr, 450 U.S. 175, 192 n.14, 209 USPQ 1, 10 n. 14 (1981). Thus, limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception do not integrate a judicial exception into a practical application.
Step 2B
Regarding the “wherein the model training request is an SQL statement” limitation, such limitation amounts to no more than generally linking the use of a judicial exception to a particular technological environment or field of use as explained above, which does not amount to significantly more than the judicial exception. MPEP 2106.05(h).
Regarding Claim 13
Step 2A, Prong 1
Claim 13 recites an apparatus that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 1 with respect to claim 1 also applies to this claim 13. While claim 13 recites additional generic computing components (“processor”, “memory”, and “instructions”), such additional generic computing components do not change the analysis under Step 2A, Prong 1.
Step 2A, Prong 2
Claim 13 recites an apparatus that corresponds to the method of claim 1, and therefore the analysis under Step 2A, Prong 2 with respect to claim 1 also applies to this claim 13. While claim 13 recites additional generic computing components (“processor”, “memory”, and “instructions”), such additional generic computing components do not change the analysis under Step 2A, Prong 2.
Step 2B
Claim 13 recites an apparatus that corresponds to the method of claim 1, and therefore the analysis under Step 2B with respect to claim 1 also applies to this claim 13. While claim 13 recites additional generic computing components (“processor”, “memory”, and “instructions”), such additional generic computing components do not change the analysis under Step 2B.
Claim 14 depends from claim 13 and claims an apparatus that corresponds to the method of claim 2 and is therefore rejected for the same reasons explained above with respect to claims 2 and 13.
Regarding Claim 17
Step 2A, prong 1 (Is the claim directed to a law of nature, a natural phenomenon or an abstract idea).
Claim 17 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including an observation, evaluation, judgment, opinion) or with the aid of pencil and paper but for the recitation of generic computer components (“non-transitory computer-readable storage medium”, “data processing apparatus”, and “instructions”).
generate an execution plan of the model training policy and an estimated execution cost of the execution plan executed by the data processing apparatus; (under the broadest reasonable interpretation, a human can mentally determine a plan for training a machine learning model, such as a plan to train a model for a particular task, using backpropagation with SGD, and a human can mentally estimate the resource cost of such plan (e.g., needing 12 hours to train on available servers))
obtain, based on the estimated execution cost, M training sample groups in the N training sample groups; (under the broadest reasonable interpretation, a human can mentally determine a subset of the training samples to use for training based on the estimated execution cost that the human wants to meet, where a human can mentally determine M, as a subset of the entire training set)
Step 2A, prong 2 (Does the claim recite additional elements that integrate the judicial exception into a practical application?).
The judicial exception is not integrated into a practical application.
Regarding the “A non-transitory computer-readable storage medium, storing a computer program including instructions that, when executed by a data processing apparatus, cause the data processing apparatus” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional elements of a non-transitory computer-readable storage medium, computer program, instructions, and data processing apparatus. These additional elements are recited at a high-level of generality and amount to no more than mere instructions to apply the exception using generic computer components. Accordingly, these additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “obtain a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups” limitation, such additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process (see MPEP 2106.05(g)).
Regarding the “executed by the data processing apparatus” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of a database system. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (a database system). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “train a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of generic training of a model. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (generic training of a model). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Regarding the “update the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception. In particular, the claim only recites the additional element of generic training of a model. This additional element is recited at a high-level of generality and amounts to no more than mere instructions to apply the exception using a generic computer component (generic training of a model). Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea (See MPEP 2106.05(f)).
Accordingly, at Step 2A, prong two, after considering all claim elements individually and as an ordered combination, it is determined that the claims do not integrate the judicial exception into a practical application.
Step 2B (Does the claim recite additional elements that amount to significantly more than the judicial exception?)
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception.
Regarding the “A non-transitory computer-readable storage medium, storing a computer program including instructions that, when executed by a data processing apparatus, cause the data processing apparatus” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “obtain a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups” limitation, as discussed above, the additional element of a data gathering step is recited at a high level of generality and amounts to extra-solution activity of receiving data, i.e. pre-solution activity of gathering data for use in the claimed process. The courts have found limitations directed to obtaining information electronically, recited at a high level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Regarding the “executed by the data processing apparatus” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “training a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Regarding the “updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model” limitation, such limitation is recited at a high-level of generality and amounts to no more than adding the words “apply it” (or an equivalent) with the judicial exception, because the limitation merely provides instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea. Accordingly, this additional element does not add significantly more than the judicial exception. (See MPEP 2106.05(f)).
Accordingly, at Step 2B after considering all claim elements individually and as an ordered combination, it is determined that the claims do not integrate the judicial exception into a practical application.
Claim 18 depends from claim 17 and claims a computer-readable storage medium that corresponds to the method of claim 2 and is therefore rejected for the same reasons explained above with respect to claims 2 and 17.
Claim 19 depends from claim 17 and claims a computer-readable storage medium that corresponds to the method of claim 3 and is therefore rejected for the same reasons explained above with respect to claims 3 and 17.
Claim 20 depends from claim 19 and claims a computer-readable storage medium that corresponds to the method of claim 4 and is therefore rejected for the same reasons explained above with respect to claims 4 and 19.
Claim Rejections - 35 USC § 103
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.
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.
Claims 1-2 and 9-14, and 17-20 are rejected under 35 U.S.C. 103 as being unpatentable over US 12387132 B1, hereinafter referenced as TONG, in view of US 20200202210 A1, hereinafter referenced as KUSHNIR, and further in view of Alonso, Gustavo, et al. "doppioDB 1.0: Machine Learning inside a Relational Engine." IEEE Data Eng. Bull. 42.2 (2019): 19-31, hereinafter referenced as ALONSO (NPL2 in Applicant’s 1/3/2025 IDS).
Regarding Claim 1
TONG teaches:
A data processing method, applied to a database system, comprising: (TONG, col. 2, line 60-67: “FIG. 1 illustrates a logical block diagram of orchestration for building and executing machine learning pipelines for graph data sets, according to some embodiments. Machine learning pipeline building system 110 may implement and/or support various interfaces to handle both requests to train and build machine learning pipelines 150 on graph data sets 142 and access graph data sets 112 in graph database 140”)
obtaining a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups; (TONG, col. 3, lines 1-5: “For example, a request 150 can include a graph data set identifier 152 and execution configuration 154 (e.g., which may include requested algorithms(s)) or other applications in which a machine learning pipeline may be deployed to build and train on graph data set 142.”;
Examiner’s Note: Request 150 includes a graph data set identifier 152 (corresponding to recited “plurality of training samples” and an execution configuration 154 (corresponding to recited “model training policy” which includes the specific algorithms to build and train), where each sample within the graph data set 142 is its own group such that N = number of samples within graph data set 142)
generating an execution plan of the model training policy and an estimated execution cost of the execution plan ; (TONG, col. 3, lines 3-14: “Machine learning pipeline building system 110 may be able to generate an execution plan (e.g., according to the various techniques discussed below with regard to FIGS. 3-5 and 7-9) performing graph profiling 122 on graph data set 142 (e.g., to determine various characteristics of graph data set 142), machine learning algorithm selection 124 (e.g., to include in the requested machine learning pipeline), and plan optimization 126 (e.g., to identify performance efficiencies to reduce building and executing costs).”
TONG, col. 3, lines 15-22: “Execution plan generation 120 may then provide the execution plan 128 to execution plan orchestration 130. Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”;
TONG, col. 5, lines 52-59: “In various embodiments, graph database service 270 may store and provide access to graphs, such as graph data set 272. These graph data sets may be used to train machine learning models. As discussed in detail below with regard to FIG. 6, in some embodiments, graph database service 270 may implement techniques for performs orchestration for building and executing machine learning models for graph data sets.”
TONG, col. 14, lines 3-26: “As discussed above with regard to FIGS. 4 and 5, various possible workflows to build and train identified models may be used. As discussed in detail below with regard to FIG. 9, techniques to optimize performance of the execution plan (e.g., by identifying sharing opportunities and/or other performance efficiencies). In some embodiments, various other configurations or options that are specified for the request to train and build the machine learning pipeline(s) may be accounted for in the execution plan (e.g., including configuration information to configure hyperparameter ranges, custom evaluation metrics, or other default automation override information). For example, access to external sources of data for reference data objects and operations to process the referenced data, like documents or images, may be included as operations in the execution plan. Operations to access data sources, select a transformer to process data objects from the data sources, such as transformers like BERT for text documents and/or pre-trained neural networks (e.g., Convolutional Neural Networks (CNN) trained for images). In some embodiments, configuration information may specify constraints like maximum cost and/or training time for performing the request (e.g., a time limit and building and executing the machine learning pipeline(s)).”;
Examiner’s Note: TONG discloses generating and executing execution plan 128, according to cost optimizations and a maximum cost constraint (corresponding to recited “estimated execution cost” conservatively assuming that the maximum cost will be met), which is orchestrated by graph database service 270)
… a currently available computing resource … (TONG, col. 3, lines 17-22: “Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”)
TONG, col. 8, line 54 – col. 9, line 2: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328. For example, some resources may be provided with greater CPU resources to handle CPU intensive tasks, while other resources in graph-based model processing 215 with greater GPU resources may handle GPU intensive tasks.”)
training a to-be-trained model in parallel (TONG, col. 8, lines 38-41: “Parallelized operations, such as various feature extraction techniques like random walk samples, 2.sup.nd order neighborhood subgraphs 514, constructed adjacency matrix 516 and constructed node feature matrices 518 may be ordered at the same time.”;
TONG, col. 8, lines 54-66: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328.”
TONG, col. 15, lines 42-58: “ As indicated at 920, operations in the execution plan may be evaluated for parallelism. For example, various feature processors or other techniques that operate on raw data may have no dependencies between them and as such may be performed in parallel. As indicated by the positive exit from 920, the initial execution plan may be modified to order the identified operations in parallel, in some embodiments.”;
Examiner’s Note: TONG teaches that operations of the machine line training pipeline can be performed in parallel)
updating the to-be-trained model ..., to obtain a trained model. (TONG, col. 3, lines 22-25: “Request result 160 may provide the trained model(s) 162 (e.g., for deployment or other serving options, as discussed below with regard to FIGS. 3, 7, and 10).”;
TONG, col. 16, lines 12-20: “Deployment of machine learning pipelines trained on graph data sets may be manually requested and performed, in some embodiments. Automation of deployment, however, may offer further performance benefits in, for instance, scenarios, where a new machine learning algorithm may be automatically deployed to replace an older version of the trained machine learning algorithm, and thus provide an automatic system update for a system that uses the trained pipeline.”;
Examiner’s Note: TONG discloses that after training the model, the model is now a trained model)
However, TONG fails to explicitly teach:
... executed by the database system
wherein the execution plan comprises a plurality of AI operators, and an operator type of each of the plurality of AI operators is preconfigured on the database system, and the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators;
obtaining, based on the estimated execution cost and … of the database system, M training sample groups in the N training sample groups, wherein a value of M is negatively correlated with the estimated execution cost and positively correlated with the currently available computing resource;
... by using the M training sample groups, to obtain M pieces of parameter update data; and
... based on the M pieces of parameter update data ...
However, in a related field of endeavor (training neural networks, see para. 0001), KUSHNIR teaches and makes obvious:
obtaining, based on the estimated execution cost and a currently available computing resource …, M training sample groups in the N training sample groups; (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using a subset of the graph data set 142 of TONG as taught by KUSHNIR, where such subset is based on a predetermined training budget as in KUSHNIR, and where the teachings of TONG are applied to identify processing resources)
wherein a value of M is negatively correlated with the estimated execution cost and positively correlated with the currently available computing resource; (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using a subset of the graph data set 142 of TONG as taught by KUSHNIR, where the subset of data points needs to be reduced to meet the training budget and maximum training cost of TONG (corresponding to recited “M is negatively correlated with the estimated execution cost”), and where there can be more data points if there are more computational resources raising the budget for computations (corresponding to recited “positively correlated with the currently available computing resource” limitation))
training a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data; and (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using the subset of data points as in KUSHNIR, where for each training iteration, an input data produces an intermediate output data (so M-to-M) corresponding to the recited “parameter update data”)
updating the to-be-trained model based on the M pieces of parameter update data to obtain a trained model. (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using the subset of data points as in KUSHNIR, where for each training iteration, an input data produces an intermediate output data (so M-to-M) corresponding to the recited “parameter update data”, where such intermediate output data is used to train the model, such as through backpropagation)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG and KUSHNIR as explained above. As disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so because KUSHNIR teaches techniques for augmenting training data set to improve labeling predictions of neural networks. (para. 0024). As further disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so in order “to find and use a training set that maximizes the accuracy of the model while keeping overall size of the training data within acceptable budgetary limits.” (para. 0023).
However, TONG and KUSHNIR fail to explicitly teach:
... executed by the database system
wherein the execution plan comprises a plurality of AI operators, and an operator type of each of the plurality of AI operators is preconfigured on the database system, and the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators;
… of the database system …
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
generating an execution plan of the model training policy and an estimated execution cost of the execution plan executed by the database system (ALONSO, p. 19, section 1: “We have implemented this idea in a first prototype of doppioDB, identified here as doppioDB 1.0 to distinguish it from future versions, showing how to integrate machine learning operators into the database engine in a way that is both efficient (i.e., compute-intensive algorithms do not impose overhead on native database workloads) and effective (i.e., standard operator models and execution patterns do not need to be modified).”;
Examiner’s Note: ALONSO teaches that a database engine can integrate and execute machine learning operators; the TONG-KUSHNIR-ALONSO combination now has the database engine of ALONSO train the models of TONG)
wherein the execution plan comprises a plurality of AI operators, and an operator type of each of the plurality of AI operators is preconfigured on the database system, and the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators; (ALONSO, p. 25, section 5: “With the structure shown in Listing 2, the user specifies (1) the model name after CREATE MODEL, (2) the attributes and the label that the model should be trained on after ON, (3) the type of the ML model after WITH (e.g., support vector machine, logistic regression, decision trees, neural networks etc.), (4) the training algorithm along with the parameters after USING (e.g., SGD, ADAM etc.).”;
ALONSO, pp. 27-28, section 6: “The trained model can be obtained from any machine learning framework for decision trees such as XGBoost”;
Examiner’s Note: the broadest reasonable interpretation of “AI operators” includes “an algorithm used when an AI-related operation is performed” as set forth in para. 0017 of the instant specification, and SGD is specifically identified as a type of AI operator; ALONSO teaches that the SGD and ADAM operators are options, and p. 28, Listing 5 shows that XGBoost is also an option; the TONG-KUSHNIR-ALONSO combination now modifies the execution plan of TONG to designate the particular types of training algorithms and optimization algorithms (e.g., SGD, ADAM, XGBoost) as disclosed by ALONSO, and now modifies the execution plan of TONG to designate the particular types of training algorithms and optimization algorithms (e.g., SGD, ADAM, XGBoost) as disclosed by ALONSO, and to determine the maximum cost according to the estimated execution costs of such algorithms)
… of the database system … (ALONSO, p. 19, section 1: “We have implemented this idea in a first prototype of doppioDB, identified here as doppioDB 1.0 to distinguish it from future versions, showing how to integrate machine learning operators into the database engine in a way that is both efficient (i.e., compute-intensive algorithms do not impose overhead on native database workloads) and effective (i.e., standard operator models and execution patterns do not need to be modified).”;
Examiner’s Note: ALONSO teaches that a database engine can integrate and execute machine learning operators; the TONG-KUSHNIR-ALONSO combination now has the database engine of ALONSO train the models of TONG, and uses the teachings of TONG (col. 3, lines 17-22 and 8, line 54- col. 9, line 2) to predict the computing resources required that are part of the database system)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 2
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. However, TONG and KUSHNIR fail to explicitly teach:
wherein the parameter update data is an update gradient of the to-be-trained model.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
wherein the parameter update data is an update gradient of the to-be-trained model. (ALONSO, p. 24, section 5: “By including a stochastic gradient descent (SGD) in doppioDB, our goal is to show that the FPGA-enabled database is capable of efficiently handling iterative model-training tasks. The SGD operator enables us to train linear regression models and support vector machines (SVM) on the FPGA using relational data as input.”
Examiner’s Note: ALONSO teaches training models using stochastic gradient descent; the TONG-KUSHNIR-ALONSO combination now has the database engine of ALONSO train the models of TONG using stochastic gradient descent to update the models during training)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 3
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. However, TONG and KUSHNIR fail to explicitly teach:
wherein the execution plan comprises a plurality of AI operators, and an operator type of each of the plurality of AI operators is preconfigured on the database system.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
wherein the execution plan comprises a plurality of AI operators, and an operator type of each of the plurality of AI operators is preconfigured on the database system. (ALONSO, p. 25, section 5: “With the structure shown in Listing 2, the user specifies (1) the model name after CREATE MODEL, (2) the attributes and the label that the model should be trained on after ON, (3) the type of the ML model after WITH (e.g., support vector machine, logistic regression, decision trees, neural networks etc.), (4) the training algorithm along with the parameters after USING (e.g., SGD, ADAM etc.).”;
ALONSO, pp. 27-28, section 6: “The trained model can be obtained from any machine learning framework for decision trees such as XGBoost”;
Examiner’s Note: the broadest reasonable interpretation of “AI operators” includes “an algorithm used when an AI-related operation is performed” as set forth in para. 0017 of the instant specification, and SGD is specifically identified as a type of AI operator; ALONSO teaches that the SGD and ADAM operators are options, and p. 28, Listing 5 shows that XGBoost is also an option; the TONG-KUSHNIR-ALONSO combination now modifies the execution plan of TONG to designate the particular types of training algorithms and optimization algorithms (e.g., SGD, ADAM, XGBoost) as disclosed by ALONSO)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 4
TONG, KUSHNIR, and ALONSO teach the method of claim 3 as explained above. However, TONG and KUSHNIR fail to explicitly teach:
wherein the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
wherein the estimated execution cost of the execution plan is obtained based on estimated execution costs of the plurality of AI operators. (ALONSO, p. 25, section 5: “With the structure shown in Listing 2, the user specifies (1) the model name after CREATE MODEL, (2) the attributes and the label that the model should be trained on after ON, (3) the type of the ML model after WITH (e.g., support vector machine, logistic regression, decision trees, neural networks etc.), (4) the training algorithm along with the parameters after USING (e.g., SGD, ADAM etc.).”;
ALONSO, pp. 27-28, section 6: “The trained model can be obtained from any machine learning framework for decision trees such as XGBoost”;
Examiner’s Note: ALONSO teaches that the SGD and ADAM operators are options, and p. 28, Listing 5 shows that XGBoost is also an option; the TONG-KUSHNIR-ALONSO combination now modifies the execution plan of TONG to designate the particular types of training algorithms and optimization algorithms (e.g., SGD, ADAM, XGBoost) as disclosed by ALONSO, and to determine the maximum cost according to the estimated execution costs of such algorithms)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 6
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. However, TONG fails to explicitly teach:
wherein a value of M is negatively correlated with the estimated execution cost.
However, in a related field of endeavor (training neural networks, see para. 0001), KUSHNIR teaches and makes obvious:
wherein a value of M is negatively correlated with the estimated execution cost. (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR-ALONSO combination now trains the models of TONG using a subset of the graph data set 142 of TONG as taught by KUSHNIR, where the subset of data points needs to be reduced to meet the training budget and maximum training cost of TONG (corresponding to recited “M is negatively correlated with the estimated execution cost”))
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so because KUSHNIR teaches techniques for augmenting training data set to improve labeling predictions of neural networks. (para. 0024). As further disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so in order “to find and use a training set that maximizes the accuracy of the model while keeping overall size of the training data within acceptable budgetary limits.” (para. 0023).
Regarding Claim 7
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. TONG further teaches:
obtaining a currently available computing resource of the database system; (TONG, col. 3, lines 17-22: “Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”)
TONG, col. 8, line 54 – col. 9, line 2: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328. For example, some resources may be provided with greater CPU resources to handle CPU intensive tasks, while other resources in graph-based model processing 215 with greater GPU resources may handle GPU intensive tasks.”)
However, TONG fails to explicitly teach:
the obtaining of the M training sample groups in the N training sample groups comprises: obtaining, based on the estimated execution cost and the currently available computing resource, the M training sample groups in the N training sample groups, wherein a value of M is positively correlated with the currently available computing resource.
However, in a related field of endeavor (training neural networks, see para. 0001), KUSHNIR teaches and makes obvious:
the obtaining of the M training sample groups in the N training sample groups comprises: obtaining, based on the estimated execution cost and the currently available computing resource, the M training sample groups in the N training sample groups, wherein a value of M is positively correlated with the currently available computing resource. (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR-ALONSO combination now trains the models of TONG using the subset of data points as in KUSHNIR, where there can be more data points if there are more computational resources raising the budget for computations)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG and KUSHNIR as explained above. As disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so because KUSHNIR teaches techniques for augmenting training data set to improve labeling predictions of neural networks. (para. 0024). As further disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so in order “to find and use a training set that maximizes the accuracy of the model while keeping overall size of the training data within acceptable budgetary limits.” (para. 0023).
Regarding Claim 9
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. TONG further teaches:
wherein there are X to-be-trained models in a one-to-one correspondence to X trained models, the X trained models are different models used to implement a target task (TONG, col. 5, lines 2-8: “Machine learning service 210 may implement graph-based pipeline serving 217, which may store and deploy various machine learning pipeline trained on graph data to perform various tasks (e.g., according to the type of machine learning algorithm, such as clustering, regression, and/or classification, among others) as part of or for various applications.”;
Examiner’s Note: TONG teaches different tasks including clustering, regression, and/or classification, where 1 to-be-trained model is trained into a trained model)
obtaining the currently available computing resource and an execution cost of each of the X trained models, and (TONG, col. 3, lines 3-14: “Machine learning pipeline building system 110 may be able to generate an execution plan (e.g., according to the various techniques discussed below with regard to FIGS. 3-5 and 7-9) performing graph profiling 122 on graph data set 142 (e.g., to determine various characteristics of graph data set 142), machine learning algorithm selection 124 (e.g., to include in the requested machine learning pipeline), and plan optimization 126 (e.g., to identify performance efficiencies to reduce building and executing costs).”
TONG, col. 3, lines 17-22: “Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”)
TONG, col. 8, line 54 – col. 9, line 2: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328. For example, some resources may be provided with greater CPU resources to handle CPU intensive tasks, while other resources in graph-based model processing 215 with greater GPU resources may handle GPU intensive tasks.”)
However, TONG and KUSHNIR fail to explicitly teach:
after the updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model, the method further comprises: obtaining a model inference request indicating the target task;
determining a target model from the X trained models; and
performing model inference by using the target model.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
after the updating the to-be-trained model based on the M pieces of parameter update data, to obtain a trained model, the method further comprises: obtaining a model inference request indicating the target task; (ALONSO, p. 25, section 5: “After the model is created, we can use it for inference on new tables using the INFER keyword, passing the model name.”;
Examiner’s Note: As shown in Listing 2 on page 25, a database system can receive a (INFER (‘model_name’) FROM new_data_set)) command from a user to perform a particular inference for the task the model was created to perform; the TONG-KUSHNIR-ALONSO combination now modifies the system of TONG to receive a model inference request from a user, such as a database systems analyst)
determining a target model from the X trained models; and (ALONSO, p. 25, section 5: “After the model is created, we can use it for inference on new tables using the INFER keyword, passing the model name.”;
Examiner’s Note: As shown in Listing 2 on page 25, a database system can receive a (INFER (‘model_name’) FROM new_data_set)) command from a user identifying a particular model; the TONG-KUSHNIR-ALONSO combination now modifies the system of TONG to receive a model inference request from a user, such as a database systems analyst, which determines which specific model to use for inference)
performing model inference by using the target model. (ALONSO, p. 25, section 5: “After the model is created, we can use it for inference on new tables using the INFER keyword, passing the model name.”;
Examiner’s Note: As shown in Listing 2 on page 25, a database system can receive a (INFER (‘model_name’) FROM new_data_set)) command from a user identifying a particular model; the TONG-KUSHNIR-ALONSO combination now modifies the system of TONG to receive a model inference request from a user, and to actually execute the INFER command)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 10
TONG, KUSHNIR, and ALONSO teach the method of claim 9 as explained above. TONG further teaches:
wherein the currently available computing resource comprises: an input/output (I/O) resource, a central processing unit (CPU) resource, a graphics processing unit (GPU) resource, and/or a memory resource. (TONG, col. 3, lines 17-22: “Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”)
TONG, col. 8, line 54 – col. 9, line 2: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328. For example, some resources may be provided with greater CPU resources to handle CPU intensive tasks, while other resources in graph-based model processing 215 with greater GPU resources may handle GPU intensive tasks.”)
Regarding Claim 11
TONG, KUSHNIR, and ALONSO teach the method of claim 9 as explained above. However, TONG and KUSHNIR fail to explicitly teach:
wherein the model inference request is a structured query language (SQL) statement.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
wherein the model inference request is a structured query language (SQL) statement. (ALONSO, p. 25, Listing 2: “
PNG
media_image6.png
258
786
media_image6.png
Greyscale
Examiner’s Note: Listing 2 shows the SELECT INFER command in SQL)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 12
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. However, TONG and KUSHNIR fail to explicitly teach:
wherein the model training request is an SQL statement.
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
wherein the model training request is an SQL statement. (ALONSO, p. 25, Listing 2: “
PNG
media_image6.png
258
786
media_image6.png
Greyscale
Examiner’s Note: Listing 2 shows the CREATE MODEL command in SQL, which includes the training algorithm and training set for training)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Regarding Claim 13
TONG teaches:
A data processing apparatus, comprising at least one processor, a memory storing instructions that, when executed by the at least one processor, cause the data processing apparatus to perform operations comprising: (TONG, col. 16, lines 54-63: “ For example, in one embodiment, the methods may be implemented on or across one or more computer systems (e.g., a computer system as in FIG. 11) that includes one or more processors executing program instructions stored on one or more computer-readable storage media coupled to the processors. The program instructions may implement the functionality described herein (e.g., the functionality of various servers and other components that implement the network-based virtual computing resource provider described herein).”)
The remaining limitations of claim 13 correspond to the method of claim 1 and therefore claim 13 is rejected for the same reasons explained above with respect to claim 1.
Claim 14 depends from claim 13 and claims an apparatus that corresponds to the method of claim 2 and is therefore rejected for the same reasons explained above with respect to claims 2 and 13.
Regarding Claim 17
TONG teaches:
A non-transitory computer-readable storage medium, storing a computer program including instructions that, when executed by a data processing apparatus, cause the data processing apparatus to: (TONG, col. 2, line 60-67: “FIG. 1 illustrates a logical block diagram of orchestration for building and executing machine learning pipelines for graph data sets, according to some embodiments. Machine learning pipeline building system 110 may implement and/or support various interfaces to handle both requests to train and build machine learning pipelines 150 on graph data sets 142 and access graph data sets 112 in graph database 140”;
TONG, col. 16, lines 54-63: “ For example, in one embodiment, the methods may be implemented on or across one or more computer systems (e.g., a computer system as in FIG. 11) that includes one or more processors executing program instructions stored on one or more computer-readable storage media coupled to the processors. The program instructions may implement the functionality described herein (e.g., the functionality of various servers and other components that implement the network-based virtual computing resource provider described herein).”)
obtain a model training request, wherein the model training request comprises a plurality of training samples and a model training policy, and the plurality of training samples is grouped into N training sample groups; (TONG, col. 3, lines 1-5: “For example, a request 150 can include a graph data set identifier 152 and execution configuration 154 (e.g., which may include requested algorithms(s)) or other applications in which a machine learning pipeline may be deployed to build and train on graph data set 142.”;
Examiner’s Note: Request 150 includes a graph data set identifier 152 (corresponding to recited “plurality of training samples” and an execution configuration 154 (corresponding to recited “model training policy” which includes the specific algorithms to build and train), where each sample within the graph data set 142 is its own group such that N = number of samples within graph data set 142)
generate an execution plan of the model training policy and an estimated execution cost of the execution plan…; (TONG, col. 3, lines 3-14: “Machine learning pipeline building system 110 may be able to generate an execution plan (e.g., according to the various techniques discussed below with regard to FIGS. 3-5 and 7-9) performing graph profiling 122 on graph data set 142 (e.g., to determine various characteristics of graph data set 142), machine learning algorithm selection 124 (e.g., to include in the requested machine learning pipeline), and plan optimization 126 (e.g., to identify performance efficiencies to reduce building and executing costs).”
TONG, col. 3, lines 15-22: “Execution plan generation 120 may then provide the execution plan 128 to execution plan orchestration 130. Execution plan orchestration 130 may identify and instruct various processing resources (e.g., as discussed below with regard to FIGS. 3 and 7), which build and train machine learning pipelines efficiently according to execution plan 128, including accessing 112 graph data set 142 in graph database 140.”;
TONG, col. 5, lines 52-59: “In various embodiments, graph database service 270 may store and provide access to graphs, such as graph data set 272. These graph data sets may be used to train machine learning models. As discussed in detail below with regard to FIG. 6, in some embodiments, graph database service 270 may implement techniques for performs orchestration for building and executing machine learning models for graph data sets.”
TONG, col. 14, lines 3-26: “As discussed above with regard to FIGS. 4 and 5, various possible workflows to build and train identified models may be used. As discussed in detail below with regard to FIG. 9, techniques to optimize performance of the execution plan (e.g., by identifying sharing opportunities and/or other performance efficiencies). In some embodiments, various other configurations or options that are specified for the request to train and build the machine learning pipeline(s) may be accounted for in the execution plan (e.g., including configuration information to configure hyperparameter ranges, custom evaluation metrics, or other default automation override information). For example, access to external sources of data for reference data objects and operations to process the referenced data, like documents or images, may be included as operations in the execution plan. Operations to access data sources, select a transformer to process data objects from the data sources, such as transformers like BERT for text documents and/or pre-trained neural networks (e.g., Convolutional Neural Networks (CNN) trained for images). In some embodiments, configuration information may specify constraints like maximum cost and/or training time for performing the request (e.g., a time limit and building and executing the machine learning pipeline(s)).”;
Examiner’s Note: TONG discloses generating and executing execution plan 128, according to cost optimizations and a maximum cost constraint (corresponding to recited “estimated execution cost” conservatively assuming that the maximum cost will be met), which is orchestrated by graph database service 270)
train a to-be-trained model in parallel (TONG, col. 8, lines 38-41: “Parallelized operations, such as various feature extraction techniques like random walk samples, 2.sup.nd order neighborhood subgraphs 514, constructed adjacency matrix 516 and constructed node feature matrices 518 may be ordered at the same time.”;
TONG, col. 8, lines 54-66: “Turning back to FIG. 3, execution plan generator 314 may provide the optimized execution plan to execution plan coordinator 316. Execution plan coordinator may direct various execution tasks, utilizing graph-based pipeline processing 214 directly and/or utilizing training tuner 318. For example, graph-based pipeline processing 215 may implement various different resources (e.g., different types of nodes or other computing systems 2000 discussed below with regard to FIG. 11), which may have installed and configured the various software applications and hardware to perform feature processing tasks 322, batch inference tasks 324, training jobs 326, and/or data integration tasks 328.”
TONG, col. 15, lines 42-58: “ As indicated at 920, operations in the execution plan may be evaluated for parallelism. For example, various feature processors or other techniques that operate on raw data may have no dependencies between them and as such may be performed in parallel. As indicated by the positive exit from 920, the initial execution plan may be modified to order the identified operations in parallel, in some embodiments.”;
Examiner’s Note: TONG teaches that operations of the machine line training pipeline can be performed in parallel)
update the to-be-trained model ..., to obtain a trained model. (TONG, col. 3, lines 22-25: “Request result 160 may provide the trained model(s) 162 (e.g., for deployment or other serving options, as discussed below with regard to FIGS. 3, 7, and 10).”;
TONG, col. 16, lines 12-20: “Deployment of machine learning pipelines trained on graph data sets may be manually requested and performed, in some embodiments. Automation of deployment, however, may offer further performance benefits in, for instance, scenarios, where a new machine learning algorithm may be automatically deployed to replace an older version of the trained machine learning algorithm, and thus provide an automatic system update for a system that uses the trained pipeline.”;
Examiner’s Note: TONG discloses that after training the model, the model is now a trained model)
However, TONG fails to explicitly teach:
... executed by the data processing apparatus
obtain, based on the estimated execution cost, M training sample groups in the N training sample groups
... by using the M training sample groups, to obtain M pieces of parameter update data; and
... based on the M pieces of parameter update data ...
However, in a related field of endeavor (training neural networks, see para. 0001), KUSHNIR teaches and makes obvious:
obtain, based on the estimated execution cost, M training sample groups in the N training sample groups; (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using a subset of the graph data set 142 of TONG as taught by KUSHNIR, where such subset is based on a predetermined training budget as in KUSHNIR, and where the teachings of TONG are applied to identify processing resources)
train a to-be-trained model in parallel by using the M training sample groups, to obtain M pieces of parameter update data; and (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using the subset of data points as in KUSHNIR, where for each training iteration, an input data produces an intermediate output data (so M-to-M) corresponding to the recited “parameter update data”)
update the to-be-trained model based on the M pieces of parameter update data to obtain a trained model. (KUSHNIR, para. 0036: “In step 206, the query unit 114 dynamically selects, from the data points in subset A determined in step 205, one or more candidate data points, which when included in the training data set along with its predicted label, most reduces or minimizes the computed entropy value of the labels predicted by the Neural Network 104 for the test data set 110. Thus, for every data point in set subset A, the data point is applied to the Neural Network unit and the entropy value of test set 110 is computed based on the labels predicted by the neural network unit for the data points in subset A. The data point(s) in A that most reduces the entropy of the test set 110 is identified as a candidate data point for addition to the training data set 108. The number candidate data points that are selected may be based on a predetermined training label budget.”;
Examiner’s Note: KUSHNIR teaches that a subset of data points (corresponding to recited “M training sample groups”) can be used for neural network training, where the number of the subset of data points is based on a predetermined training label budget; the TONG-KUSHNIR combination now trains the models of TONG using the subset of data points as in KUSHNIR, where for each training iteration, an input data produces an intermediate output data (so M-to-M) corresponding to the recited “parameter update data”, where such intermediate output data is used to train the model, such as through backpropagation)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG and KUSHNIR as explained above. As disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so because KUSHNIR teaches techniques for augmenting training data set to improve labeling predictions of neural networks. (para. 0024). As further disclosed by KUSHNIR, one of ordinary skill would have been motivated to do so in order “to find and use a training set that maximizes the accuracy of the model while keeping overall size of the training data within acceptable budgetary limits.” (para. 0023).
However, TONG and KUSHNIR fail to explicitly teach:
... executed by the data processing apparatus
However, in a related field of endeavor (machine learning with respect to database systems, see p. 19, section 1), ALONSO teaches and makes obvious:
generating an execution plan of the model training policy and an estimated execution cost of the execution plan executed by the data processing apparatus (ALONSO, p. 19, section 1: “We have implemented this idea in a first prototype of doppioDB, identified here as doppioDB 1.0 to distinguish it from future versions, showing how to integrate machine learning operators into the database engine in a way that is both efficient (i.e., compute-intensive algorithms do not impose overhead on native database workloads) and effective (i.e., standard operator models and execution patterns do not need to be modified).”;
Examiner’s Note: ALONSO teaches that a database engine can integrate and execute machine learning operators; the TONG-KUSHNIR-ALONSO combination now has the database engine of ALONSO train the models of TONG)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, and ALONSO as explained above. As disclosed by ALONSO, one of ordinary skill would have been motivated to do so because ALONSO teaches that it is “important to extend the role of the database management system to a more comprehensive platform supporting complex and computationally intensive data processing.” (p. 19, section 1). As further disclosed by ALONSO, one of ordinary skill would have been motivated to do so because “businesses have massive amounts of data in their existing DBMS and there is a high potential for using ML to extract valuable information from it. In addition, the rich relational operators provided by the DBMS can be used conveniently to denormalize a complex schema for the purposes of ML tasks.” (p. 20, section 1).
Claim 18 depends from claim 17 and claims a computer-readable storage medium that corresponds to the method of claim 2 and is therefore rejected for the same reasons explained above with respect to claims 2 and 17.
Claim 19 depends from claim 17 and claims a computer-readable storage medium that corresponds to the method of claim 3 and is therefore rejected for the same reasons explained above with respect to claims 3 and 17.
Claim 20 depends from claim 19 and claims a computer-readable storage medium that corresponds to the method of claim 4 and is therefore rejected for the same reasons explained above with respect to claims 4 and 19.
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over TONG in view of KUSHNIR and ALONSO and further in view of US 20130086038 A1, hereinafter referenced as PERRY.
Regarding Claim 5
TONG, KUSHNIR, and ALONSO teach the method of claim 4 as explained above. However, TONG, KUSHNIR, and ALONSO fail to explicitly teach:
Obtaining the estimated execution costs of the plurality of AI operators by using an execution plan query statement.
However, in a related field of endeavor (accessing data in databases using structured query language, see para. 0002), PERRY teaches and makes obvious:
obtaining, based on the plurality of AI operators, the estimated execution costs of the plurality of AI operators by using an execution plan query statement. (PERRY, para. 0068: “Once all SQL statements of the workload have been selected, and have had associated estimation/execution costs stored within the explain data 324, then the candidate index provider 328 may proceed by fetching the stored explain/cost data (420).”;
Examiner’s Note: the TONG-KUSHNIR-ALONSO-PERRY combination now uses the “explain” data 324, accessed by a SQL command, to store cost data of TONG)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, ALONSO, and PERRY as explained above. As disclosed by PERRY, one of ordinary skill would have been motivated to do so to provide “a user with an ability to easily design and implement database queries in a manner which returns fast and accurate query results” using the machine learning techniques of TONG as applied to database systems as in ALONSO. (para. 0009).
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over TONG in view of KUSHNIR and ALONSO and further in view of US 20210073909 A1, hereinafter referenced as LE ROUX.
Regarding Claim 8
TONG, KUSHNIR, and ALONSO teach the method of claim 1 as explained above. However, TONG, KUSHNIR, and ALONSO fail to explicitly teach:
performing a shuffle operation on the plurality of training samples, to obtain a plurality of shuffled training samples; and grouping the plurality of shuffled training samples, to obtain the N training sample groups.
However, in a related field of endeavor (training machine learning models, see para. 0004), LE ROUX teaches and makes obvious:
performing a shuffle operation on the plurality of training samples, to obtain a plurality of shuffled training samples; and grouping the plurality of shuffled training samples, to obtain the N training sample groups. (LE ROUX, para. 0109: “In one embodiment, the training data, which may be obtained from the various sources, is shuffled or otherwise processed to increase the speed of the training, reduce variance, and/or reduce overfitting.”;
Examiner’s Note: the TONG-KUSHNI-ALONSO-LE ROUX combination now shuffles the training data of TONG as taught by LU, and then re-groups the entire training set (where each group is a single sample) to obtain the N training sample groups, just in a shuffled order as in LE ROUX)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the teachings of TONG, KUSHNIR, ALONSO, and LE ROUX as explained above. As disclosed by LE ROUX, one of ordinary skill would have been motivated to do so to “increase the speed of the training, reduce variance, and/or reduce overfitting.” (para. 0109).
Conclusion
THIS ACTION IS MADE FINAL. 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 MICHAEL C LEE whose telephone number is (571)272-4933. The examiner can normally be reached M-F 12:00 pm - 8:00 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, Omar Fernandez Rivas can be reached at 571-272-2589. 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.
/MICHAEL C. LEE/Examiner, Art Unit 2128
Prosecution Insights