DETAILED ACTION
Status of Claims
Claim(s) 1-20 are pending and are examined herein.
Claim(s) 1, 11, and 20 have been Amended.
Claim(s) 1-20 are rejected under 35 U.S.C. § § 101 and 103.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 12/08/2025 has been entered.
Response to Arguments
Applicant's arguments with respect to the rejection under 35 U.S.C. § 101, filed on 12/08/2025, have been fully considered but are not persuasive. (See remark p. 8).
The Applicant submits that the claims 1, 11, and 20 have been amended and that these amendments render the § 101 rejection of this claim moot.
The Examiner notes that these amendments fail to provide the technical implementation (additional elements) that would integrate the abstract idea into a practical application and/or amount to significantly more. Specifically, Applicant’s amendment does not provide specific technical details regarding the bias detection mechanism, how the bias lineage mitigates or eliminates bias in the federated learning process, or how the tree-based bias lineage data structure is technically generated. Thus, the amendment does not provide sufficient additional elements to integrate the judicial exception into a practical application or provide significantly more.
Applicant's arguments, with respect to the rejection under 35 U.S.C. § 103 filed on 12/08/2025 (see remark Pp. 8-9) have been fully considered but are moot in view of the new ground of rejection.
The examiner refers to the updated rejection under 35 U.S.C. § 103 for more details.
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.
When considering subject matter eligibility under 35 U.S.C. 101, it must be determined whether the claim is directed to one of the four statutory categories of invention, i.e., process, machine, manufacture, or composition of matter (Step 1). If the claim does fall within one of the statutory categories, the second step in the analysis is to determine whether the claim is directed to a judicial exception (Step 2A). The Step 2A analysis is broken into two prongs. In the first prong (Step 2A, Prong 1), it is determined whether or not the claims recite a judicial exception (e.g., mathematical concepts, mental processes, certain methods of organizing human activity). If it is determined in Step 2A, Prong 1 that the claims recite a judicial exception, the analysis proceeds to the second prong (Step 2A, Prong 2), where it is determined whether or not the claims integrate the judicial exception into a practical application. If it is determined at step 2A, Prong 2 that the claims do not integrate the judicial exception into a practical application, the analysis proceeds to determining whether the claim is a patent-eligible application of the exception (Step 2B). If an abstract idea is present in the claim, any element or combination of elements in the claim must be sufficient to ensure that the claim integrates the judicial exception into a practical application, or else amounts to significantly more than the abstract idea itself. Applicant is advised to consult MPEP 2106 for more details of the analysis.
Under Step 1 analysis,
Claims 1-10 recite a method (representing a process); Claims 11-19 recite an apparatus (representing a machine); and Claim 20 recites a non-transitory computer medium (representing an article of manufacture).
Therefore, each set of the claims falls into one of the four statutory categories (i.e., process, machine, article of manufacture, or composition of matter).
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (i.e., a law of nature, a natural phenomenon, or an abstract idea) without significantly more, and hence is not patent-eligible subject matter.
Step 2 Analysis (Combined Step 2A Prong 1-2 and Step 2B Analysis)
Claim 1 includes the following recitation of an abstract idea:
constructing … based on the bias metrics, bias lineages for the aggregated machine learning models using a tree-shaped data structure where each node represents a model version and parent-child relationships indicate version derivation; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. Examiner notes: the “constructing” step, as currently drafted, covers a process that can be practically performed in the human mind with the aid of a pen and paper. Constructing a lineage tree graph representing models version can be performed manually with pen and paper.)
Claim 1 recites the following additional elements which, considered individually and as an ordered combination, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
by the device (This is a high-level recitation of generic computer components for performing the abstract idea. This does not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea. See MPEP 2106.05(f).)
receiving, at a device and from a plurality of training nodes that train a set of machine learning models using local training datasets, bias metrics associated with those machine learning models for each feature of the local training datasets (This is insignificant extra-solution activity. See MPEP 2106.05(g). Moreover, sending or receiving data is well-understood, routine, conventional as evidenced by the court cases cited at MPEP 2106.05(d), example i. Receiving or transmitting data.),
wherein the bias metrics are indicative of recall and precision, the bias metrics comprising confusion matrices for sub-populations corresponding to each feature; (This limitation is an additional element that describes the content of the bias metrics (i.e., confusion matrices ) as including data indicative of recall and precision. This is a limitation that amount to merely indicating a field of use or technological environment in which to apply a judicial exception and does not amount to significantly more than the exception itself, see MPEP 2106.05(h));
generating, by the device, aggregated machine learning models over time that aggregate the machine learning models trained by the plurality of training nodes using federated learning techniques; (This is an attempt to limit the abstract idea to a particular field of use or technological environment, which does not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea. See MPEP 2106.05(h). Examiner notes that the high-level recitation of “using federated learning techniques” merely represents a technological environment and/or adding insignificant extra-solution activity to the judicial exception, as discussed in MPEP § 2106.05(g). The concept of using federated learning techniques to aggregate ML models represents a generic/conventional function in the art. As evidenced by Greenewald et al., (Pub. No: US 20210089878 A1), which explicitly states the traditional/common approach in federated learning as training local models, sending parameters to a server to generate the global model (see Greenewald [0038] and [0085]).Accordingly, this additional element does not add meaningful limitation to the process as it represents well-understood, routine, conventional activity in the field.
providing, by the device and based on the bias lineages, a bias lineage for a particular one of the aggregated machine learning models for display (This is insignificant extra-solution activity. See MPEP 2106.05(g). Moreover, sending or receiving data is well-understood, routine, conventional as evidenced by the court cases cited at MPEP 2106.05(d), example i. Receiving or transmitting data.)
Accordingly, when viewed as a whole, the claim is primarily directed to the abstract idea of analyzing information bias metrics and its associated content and constructing a bias lineage tree-shaped data structure (i.e., graph), representing information (i.e., model versions derivation), and displaying the result of the bias linage. The additional limitations, whether considered individually or in combination with the judicial exception, merely define pre-solution and post-solution activities incidental to the primary process and constitute a nominal or tangential addition and/or a high-level recitation of generic computing components.
Therefore, the claim fails to provide sufficient additional technical elements that would integrate the abstract idea into a practical application or amount to significantly more.
Claim 2 does not reflect an improvement to computer technology or any other technology.
Claim 3 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 3 recites the following additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
receiving, from the plurality of training nodes, parameters for the set of machine learning models, wherein the device generates the aggregated machine learning models based on those parameters (This is insignificant extra-solution activity. See MPEP 2106.05(g). Moreover, sending or receiving data is well-understood, routine, conventional as evidenced by the court cases cited at MPEP 2106.05(d), example i. Receiving or transmitting data.) Accordingly, claim 3 does not reflect an improvement to computer technology or any other technology.
Claim 4 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 4 includes the following recitation of an additional abstract idea:
the bias lineage for the particular one of the aggregated machine learning models indicates at least one of the local training datasets as a source of bias for a data feature used by that aggregated machine learning model (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. This is also a part of the abstract idea identified above and is consequently not an additional element for the purposes of the 2B analysis.) Claim 4 recites no further additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, that integrate the abstract idea into a practical application or amount to significantly more than the abstract ideas. Accordingly, claim 4 does not reflect an improvement to computer technology or any other technology.
Claim 5 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 5 recites the following additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
the plurality of training nodes do not transfer their local training datasets to the device (This appears to be an attempt to limit data to a particular type or source, which amounts to no more than an attempt to limit the abstract idea to a particular field of use or technological environment. This does not provide the significantly more element required to overcome the abstract idea. See MPEP 2106.05(h).) Accordingly, claim 5 does not reflect an improvement to computer technology or any other technology.
Claim 6 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 6 includes the following recitation of an additional abstract idea:
the bias lineage for the particular one of the aggregated machine learning models indicates a version of at least one machine learning model on which it is based (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. This is also a part of the abstract idea identified above and is consequently not an additional element for the purposes of the 2B analysis.) Claim 6 recites no further additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, that integrate the abstract idea into a practical application or amount to significantly more than the abstract ideas. Accordingly, claim 6 does not reflect an improvement to computer technology or any other technology.
Claim 7 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 7 includes the following recitation of an additional abstract idea:
the bias lineage further indicates which of the plurality of training nodes associated with the at least one machine learning model on which the particular one of the aggregated machine learning models is based (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. This is also a part of the abstract idea identified above and is consequently not an additional element for the purposes of the 2B analysis.) Claim 7 recites no further additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, that integrate the abstract idea into a practical application or amount to significantly more than the abstract ideas. Accordingly, claim 7 does not reflect an improvement to computer technology or any other technology.
Claim 8 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 8 includes the following recitation of an additional abstract idea:
excluding a machine learning model from a particular one of the plurality of training nodes from being used to generate one of the aggregated machine learning models, based on a determination that the bias metrics from that particular training node exceed a threshold (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. The threshold determination portion is also a mathematical concept.) Claim 8 recites no further additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, that integrate the abstract idea into a practical application or amount to significantly more than the abstract ideas. Accordingly, Claim 8 does not reflect an improvement to computer technology or any other technology.
Claim 9 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 9 includes the following recitation of an additional abstract idea:
the bias lineage indicates that the machine learning model from the particular one of the plurality of training nodes was excluded from being used to generate one of the aggregated machine learning models (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components. This is also a part of the abstract idea identified above and is consequently not an additional element for the purposes of the 2B analysis.) Claim 9 recites no further additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, that integrate the abstract idea into a practical application or amount to significantly more than the abstract ideas. Accordingly, claim 9 does not reflect an improvement to computer technology or any other technology.
Claim 10 recites at least the abstract idea identified above in the claim upon which it depends.
Claim 10 recites the following additional elements which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
rolling back the particular one of the aggregated machine learning models to a prior version … based in part on the bias lineage for the particular one of the aggregated machine learning models (This falls under mere instructions to apply the bias lineage(s) in order to roll-back as prior version(s) are included in a bias lineage and would enable the feature to use past versions of models. See MPEP 2106.05(f).)
in response to a request to do so from a user interface (This is insignificant extra-solution activity. See MPEP 2106.05(g). Moreover, sending or receiving data is well-understood, routine, conventional as evidenced by the court cases cited at MPEP 2106.05(d), example i. Receiving or transmitting data.) Accordingly, Claim 10 does not reflect an improvement to computer technology or any other technology.
Claim 11 recites at least the abstract idea identified above in claim 1.
Claim 11 recites the following additional elements aside from those described in claim 1 which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
one or more network interfaces; a processor coupled to the one or more network interfaces and configured to execute one or more processes; and a memory configured to store a process that is executable by the processor (This is a high-level recitation of generic computer components for performing the abstract idea. This does not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea. See MPEP 2106.05(f).) Accordingly, claim 11 does not reflect an improvement to computer technology or any other technology.
Claims 12-19 recite at least the abstract idea identified above in the claim upon which they depend.
Claims 12-19 recite substantially similar subject matter to claims 2-9, respectively, and are rejected with the same rationale, mutatis mutandis. Claims 12-19 do not reflect an improvement to computer technology or any other technology.
Claim 20 recites at least the abstract idea identified above in claim 1.
Claim 20 recites the following additional elements aside from those described in claim 1 which, considered individually and as an ordered combination with the additional elements from the claim upon which it depends, do not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea:
tangible, non-transitory, computer-readable medium storing program instructions that cause a device to execute a process (This is a high-level recitation of generic computer components for performing the abstract idea. This does not integrate the abstract idea into a practical application or amount to significantly more than the abstract idea. See MPEP 2106.05(f).) Accordingly, claim 20 does not reflect an improvement to computer technology or any other technology.
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.
Claim(s) 1-3, 5-8, 11-13, 15-18, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Watanabe et al., (Pub. No.: US 20230050708 A1) in view of Das et al., (Pub. No.: US 20220171991 A1), and further in view of Thomas et al., (Pub. No.: US 11586847 B1). Hereinafter, the combination of Watanabe, Das, and Thomas.
Regarding Currently Amended Claim 1,
Watanabe discloses the following:
A method comprising: (Watanabe, [0049] “a method 200 of training a federated learning model in accordance with an embodiment of the present invention.” [0071] “FIG. 7 is a block diagram depicting components of a computer 10 suitable for executing the methods disclosed herein. Computer 10 may implement central training server 105 and/or distributed computing nodes 130A-130N in accordance with embodiments of the present invention.”)
receiving, at a device and from a plurality of training nodes that train a set of machine learning models using local training datasets (Watanabe, [0024]-[0025] “FIG. 1 is a block diagram depicting a computing environment 100 for training a federated learning model in accordance with an embodiment of the present invention. As depicted, computing environment 100 includes a central training server 105, distributed computing nodes 130A-130N, and a network 160… central training server 105 oversees federated learning tasks, analyzes distributed computing nodes to identify any nodes having biased local training data sets, and performs actions to substantially reduce or eliminate bias during training tasks.” [0030] “Node analysis module 120 receives or obtains statistical data from computing nodes participating in a federated learning task, and analyzes the data to identify any computing nodes whose local training data is biased beyond a threshold amount.” [0054] “The results of training are received from the computing nodes and a trained global model is generated at operation 250. The results of training the distributed model, such as the weight values and bias values of a neural network, may be received by central training server 105.”), bias metrics associated with those machine learning models for each feature of the local training datasets (Watanabe, [0030]-[0031] “In particular, node analysis module 120 identifies bias in local training data of distributed computing nodes 130A-130N when it is determined that some types of data samples are overrepresented or underrepresented with respect to other types of samples… the statistical data associated with a distributed computing node may include a count of data samples for each of the five labels.” [0032] “node analysis module 120 identifies bias in local training data sets by comparing the count of data samples for an overrepresented label to the count of data samples for other labels.”), … ;
generating, by the device, aggregated machine learning models over time that aggregate the machine learning models trained by the plurality of training nodes using federated learning techniques (Watanabe, [0029] “Federated learning module 115 may combine the results from multiple computing nodes to generate a global trained model, which can be generated using any conventional or other federated learning techniques. In general, a global model may be generated based on parameters of the trained distributed models, such as the weights and biases of a trained distributed model's neural network; values for the global model can be obtained by averaging or otherwise combining corresponding values of multiple trained distributed models.” [0054] “The results of training are received from the computing nodes and a trained global model is generated at operation 250. The results of training the distributed model, such as the weight values and bias values of a neural network, may be received by central training server 105… The trained global model may be produced by combining the results of training received by distributed computing nodes 130A-103N according to any conventional or other federated learning techniques.” Further see [0027].) [Examiner’s Note: This defines aggregation of trained distributed models into a global model via federated learning, performed on iterative FL rounds.];
…..
providing, by the device …. [information] for a particular one of the aggregated machine learning models for display (Watanabe, [0078] “I/O interface(s) 22. I/O interface(s) 22 may also connect to a display 30. Display 30 provides a mechanism to display data to a user and may be, for example, a computer monitor.” [0082] “The present invention embodiments may employ any number of any type of user interface (e.g., Graphical User Interface (GUI), command-line, prompt, etc.) for obtaining or providing information (e.g., data relating to training a federated learning model), where the interface may include any information arranged in any fashion. The interface may include any number of any types of input or actuation mechanisms (e.g., buttons, icons, fields, boxes, links, etc.) disposed at any locations to enter/display information and initiate desired actions via any suitable input devices (e.g., mouse, keyboard, etc.). The interface screens may include any suitable actuators (e.g., links, tabs, etc.) to navigate between the screens in any fashion.”) [Examiner’s Note: Watanabe discloses GUI/display mechanisms and component for providing information related to trained federated learning models to a user].
Watanabe is salient and does not appear to explicitly recite that:
wherein the bias metrics are indicative of recall and precision, the bias metrics comprising confusion matrices for sub-populations corresponding to each feature; and
constructing, by the device and based on the bias metrics, bias lineages for the aggregated machine learning models using a tree-shaped data structure where each node represents a model version and parent-child relationships indicate version derivation; and
The displayed information being “based on the bias lineages” and representing “a bias lineage” (i.e., the content and representation of the display).
However, Watanabe in view of Das teaches the following limitations. Hereinafter, Das teaches:
receiving, at a device and from a plurality of training nodes that train a set of machine learning models using local training datasets (Das, [0072] “Machine learning service 210 may implement model training nodes 214 to execute training jobs on various machine learning models using data sets, such as data sets 232 in storage services 230 across one or more training nodes...” [0104]-[0108] “Each trial may have associated bias and/or feature attribution metrics as specified in training job 370 such that different views or reports, as discussed in detail below with regard to FIGS. 6-8 and 15 may be generated. Report generation 360 may be implemented to store various bias metrics for pre and post training, as well as feature attribution measurements as part of fairness report(s) 390 in storage service 230… Leader node 410 may be one of a cluster of training nodes(e.g., 310 or 214), …Distributed feature attribution measurement management 420 may assign input data set partitions to different worker nodes, as indicated at 436 a, 436 b, and 436 b. as indicated at 444 a, 444 b, and 444 c, and returned to worker nodes 430, to generate and return a feature attribution measurement, as indicated at 438 a, 438 b, and 438 c to leader node 410. Leader node 410 may combine and store the respective feature attribution measurements, as indicated at 450.” [0138] “The respective feature attribution measurements may then be returned to the leader node. As indicated at 1440, the leader node may combine the respective feature attribution measurements from the worker nodes to store as the feature attribution for the machine learning model, in some embodiments. The combined measurement may, for instance, be performed according to the specified aggregation parameter (e.g., mean, mean squared, median). The combined measurement may be stored, as discussed above with regard to FIG. 12.”), bias metrics associated with those machine learning models for each feature of the local training datasets (Das, [0025] “In a training data set, the number of labels of types 0 and 1 may be counted, grouped by the restricted feature (denoted class Xr, e.g., gender, also denoted earlier as the demographic group or attribute of interest (which may be referred to as a “facet” in some embodiments), which designates the sample into the advantaged group (favored by the bias and marked/subscripted by a) and the disadvantaged group (disfavored by bias and marked/subscripted by d).” [0081] “Bias measurement 330 may support the determination of various pre-training and post-training bias metrics as discussed above with regard to FIG. 1. For example, training job 370 may specify various features, such as the attribute of interest (or sensitive attribute Xr) 371, … selections of one or more pre-training metrics 372…” Further See [0036]-[0037].) wherein the bias metrics are indicative of recall and precision (Das, [0041]-[0043] “Recall Difference (RD): A metric may be for knowing whether there is a difference in recall of the model across the attributes of interest. …RD may be the difference in recall for the advantaged vs. disadvantaged group, described as … Precision difference (PD). If the precision for the advantaged class is greater than the precision for the disadvantaged class, it implies that more of the predicted true positives are valid compared to false positives for the advantaged class. This suggests that predictions favor the advantaged class. PD may be the difference in precision for the advantaged vs. disadvantaged classes.”), the bias metrics comprising confusion matrices for sub-populations corresponding to each feature; (Das, [0025] “In a training data set, the number of labels of types 0 and 1 may be counted, grouped by the restricted feature (denoted class Xr, e.g., gender, also denoted earlier as the demographic group or attribute of interest (which may be referred to as a “facet” in some embodiments), which designates the sample into the advantaged group (favored by the bias and marked/subscripted by a) and the disadvantaged group (disfavored by bias and marked/subscripted by d).” [0063] “RD may be the difference in recall for the advantaged vs. disadvantaged group, described as
R
D
=
T
P
a
T
P
a
+
F
N
a
-
T
P
d
T
P
d
+
F
N
d
… Two types of bias metrics for this difference may be (a) Difference in acceptance rates (DAR): This metric measures whether qualified applicants from the advantaged and disadvantaged classes are accepted at the same rates. It is the difference in the ratio of true positives to predicted positives for each class; (b) Difference in rejection rates (DRR): This metric measures whether qualified applicants from the advantaged and disadvantaged class are rejected at the same rates. It is the difference in the ratio of true negatives to predicted negatives for each class.”) [Examiner’s Note: Das teaches the confusion matrices of the bias metric measurements] the bias metrics comprising confusion matrices (e.g., TP, TN) computed for advantaged and disadvantage sub-populations (subgroups) of each specified feature/attribute.]
providing, by the device and based on the bias lineages, a bias lineage for a particular one of the aggregated machine learning models for display. (Das, [0116] “FIG. 7 illustrates an example explainability view for a trained model, according to some embodiments. Development and management environment 213 may provide a global model explainability view 710, which may show an explanation of the model 720, as well as an illustrate of the impact different features have on the model, as indicated by bar graph 730 for features 740 a, 740 b, 740 c, and 740 d.” [0122]-[0124] “monitoring job view 1030 may be generated as part of development and management environment 213. Monitoring job view 1030 may provide a display of bias metric change over time 1031, indicating the bias metric 1033, as well as other monitoring features, ... explainability configuration 963 may support monitoring to detect global model feature attribution drift. Fairness and explainability monitoring 920 may obtain various reference feature attributions, from training reports or past measurements computed by feature attribution measurement 940 (which may perform global feature attribution measurement according to the techniques discussed above with regard to FIGS… For example, monitoring job view 1050 may implement illustrate feature impact on model change over time 1051. For example, each feature, such as features 1052 a, 1052 b, 1052 c, and 1052 d, may be illustrated in order of impact.”) [Examiner’s Note: Das teaches a traceable record of bias metrics changing/evolving over time. The monitoring views (graphs) providing e.g., bias metric change over time graph, feature impact on the model change over time. The record showing how the bias metrics derived and evolved read on the “bias lineage.”]
Accordingly, at the effective filing date, it would have been prima facie obvious to one ordinarily skill in the art to modify the combination of Watanabe and Das to incorporate various methods and techniques for measurements and explainability of machine learning pipelines as taught by Das. One would have been motivated to make such a combination in order to collect, aggregate, and process large amounts of fine-grained data using cloud computing, and improve access to increasingly sophisticated machine learning models that can analyze this data. Doing so would improve accuracy, increase productivity, and cost saving (Das [0001]).
Watanabe in view of Das does not appear to explicitly teach:
constructing, by the device and based on the bias metrics, bias lineages for the aggregated machine learning models using a tree-shaped data structure where each node represents a model version and parent-child relationships indicate version derivation;
However, Thomas, in combination with Watanabe and Das, teaches the following:
constructing, by the device and based on the bias metrics, bias lineages for the aggregated machine learning models using a tree-shaped data structure where each node represents a model version and parent-child relationships indicate version derivation; (Thomas, [Col. 18, Lines 5-30] “FIG. 6 illustrates an example presentation by an experiment management service of ancestor-descendant relationships among a plurality of machine learning models used for experiment trials, … In response to a programmatic request for lineage information pertaining to one or more of the model versions (e.g., similar to the GetModelLineagelnfo request discussed in the context of FIG. 4 ), a graphical interface 601 may be used to indicate the logical relationships between the related model versions of the trials. A child or direct descendant relationship with respect to a parent model version may indicate that, for example, one or more hyper-parameters, model settings, or portions of model source code of the parent were changed to derive the child version. For example, model version 610B is shown as being derived or descended from model version 610; model versions 610C and 610D are both shown as direct descendants or children of model version 610B; and model versions 610E, 610F and 610G are shown as children of model version 610D.” [Col. 9, Lines 40-60] “The EMS 132 may enable its clients to organize and compare machine learning-related tasks using a hierarchy of entities organized into sharable workspaces called projects in the depicted embodiment… Repository 138 may comprise metadata about the entity hierarchy for various projects, including information about instances of the entities, associated artifacts, containment relationships (e.g., which experiment contains which trials), temporal relationships, as well as lineage or ancestor-descendant relationships (e.g., whether the model version used in trial T7 of an experiment is derived from the model version used in trial T2, or whether the model version used in trial T7 was derived from the model version used in trial T4), and so on. Such information may be presented to EMS clients in response to programmatic requests received via interfaces 177B.” [Col. 5, Lines 55-65; Col. 6, Lines 1-3.] “ the EMS provide an indication of one or more of: a change in model quality achieved between a first and a second version of a machine learning model utilized in the group of machine learning tasks, an indication of respective resources consumed for individual machine learning tasks of the group or ancestor-descendant relationships between at least one pair of versions of machine learning models utilized in the group of machine learning tasks. In some embodiments, an EMS may automatically capture results and performance metrics and/or quality metrics during various trials and other experiment related tasks (e.g., based on hints provided by EMS clients, or based on contents of a knowledge base about various kinds of metrics collected/utilized in earlier experiments). In one embodiment, at least some of the tasks for which metadata is stored at an EMS may include a preparation procedure such as data record labeling or cleansing/reformatting.” [Col. 24, Lines 29-32] “the artifacts themselves may be distributed across multiple services of one or more provider networks, and the EMS may act as a centralized clearinghouse or aggregator for the artifacts.” Further See [Col. 3, Lines 60-62; Col. 4, Lines 5-10].) [Examiner’s Note: Thomas teaches generating a tree-shaped (hierarchy graph) representing model versions based on lineage information, where each model version is associated with metadata including various metrics collected during experiments or trials of machine learning model tasks. The disclosed metrics of Thomas would encompass kinds of metrics for ML models. It is therefore noted that the metrics collected as taught by Thomas would include the bias metrics generated and identified previously by Das.] and
providing, by the device and based on the bias lineages, a bias lineage for a particular one of the aggregated machine learning models for display. (Thomas, [Col. 18, Lines 5-30] “FIG. 6 illustrates an example presentation by an experiment management service of ancestor-descendant relationships among a plurality of machine learning models used for experiment trials, according to at least some embodiments. In the depicted embodiment, one or more experiments managed with the help of the EMS may comprise at least seven trials 650A—650G), in each of which a respective version of a machine learning model is utilized… a graphical interface 601 may be used to indicate the logical relationships between the related model versions of the trials. A child or direct descendant relationship with respect to a parent model version may indicate that, for example, one or more hyper-parameters, model settings, or portions of model source code of the parent were changed to derive the child version. ” [Col. 18, Lines 45-60] “In the depicted embodiment, in addition to showing the ancestor-descendant relationships among model versions used in various trials, a number of other types of information may be displayed. For each of the model versions a respective timestamp 650 (e.g., 650A-650G) indicating for example when the model version was created or last modified may be shown. A respective “View details” interface element 690 (e.g., 690A-690G) may be presented in some embodiments, enabling an EMS client access additional artifacts associated with the model versions, such as notes or annotations, source code, etc. In at least one embodiment, a “View diffs” interface element 691 (e.g., 691A-691G) may be provided if a client wishes to view the specific differences between a pair of model versions (e.g., hyper-parameter differences, source code differences, performance or resource consumption differences, and so on). Other types of lineage-related information than the elements shown in FIG. 6 may be presented by the EMS in some embodiments.”)
Accordingly, it would have been obvious to a person having ordinary skill in the art, before the effective filing date of the claimed invention, having the combination of Watanabe, Das, and Thomas, to incorporate the methods and apparatus for implementing a network-accessible experiment management service as taught by Thomas. One would have been motivated to make such a combination in order to improve the overall quality of machine learning solutions by making it easier to reproduce (and thereby validate) or invalidate results obtained in earlier experiments. Doing so would reduce the total amount of resources consumed for generating high quality machine learning models and providing clear insights into the factors that result in the largest quality improvements (Thomas [Col. 5, Lines 15-25]).
Regarding Original Claim 2, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
wherein the plurality of training nodes are located across a plurality of different geographic locations. (Watanabe, [0017] “Federated learning approaches aim to address this issue, as training can be accomplished without requiring the individual training nodes to share their local sets of training data with the other training nodes or other entities. Thus, sensitive data can be safely used to contribute toward the training of a global model without compromising the privacy of the data.” [0038] “Network 160 may include a local area network (LAN), a wide area network (WAN) such as the Internet, or a combination of the two, and includes wired, wireless, or fiber optic connections. In general, network 160 can be any combination of connections and protocols known in the art that will support communications between central training server 105 and distributed computing nodes 130A-130N via their respective network interfaces in accordance with embodiments of the present invention.” [0050] “The nodes may each be associated with a different entity or organization, and as such, may each include a separate volume of data that can be used to locally train the distributed portion of a federated learning model while ensuring the data's privacy.” Further see [0086].)
wherein the plurality of training nodes are located across a plurality of different geographic locations. (Thomas, [Col. 9, Lines 10-16] “One or more training and evaluation coordinators 174 (also referred to as model generators) implemented at a set of computing devices may utilize training resources 110 (e.g., compute engines including CPUs, GPUs and the like, as well as storage, memory and networking resources) to train machine learning models using algorithms from library 120 in the depicted embodiment.” [Col. 22, Lines 5-15] “Networks set up by an entity such as a company or a public sector organization to provide one or more network-accessible services (such as various types of cloud-based computing, storage or analytics services) accessible via the Internet and/or other networks to a distributed set of clients may be termed provider networks. A provider network may sometimes be referred to as a “public cloud” environment. The resources of a provider network may in some cases be distributed across multiple data centers, which in turn may be distributed among numerous geographical regions (e.g., with each region corresponding to one or more cities, states or countries. For example, a cloud provider network can be formed as a number of regions, where a region is a geographical area in which the cloud provider clusters data centers.”)
Regarding Original Claim 3, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
wherein metadata and runtime characteristics relating to the quantum computing function are collected from the execution environment while the quantum computing function is being executed. (Watanabe, [0029] “In general, a global model may be generated based on parameters of the trained distributed models, such as the weights and biases of a trained distributed model's neural network; values for the global model can be obtained by averaging or otherwise combining corresponding values of multiple trained distributed models.” [0043] “Training module 140 may share the results of training, such as the weight values and bias values of a neural network, or other data corresponding to a trained model, with central training server 105 so that a global model may be developed.”)
Regarding Original Claim 5, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
wherein the plurality of training nodes do not transfer their local training datasets to the device. (Watanabe, [0002] “Federated learning, also known as collaborative learning, is a machine learning technique in which algorithms are trained across multiple decentralized computing nodes holding local data samples, without exchanging data between the computing nodes.” [0031] “Node analysis module 120 may analyze statistical data that describes the local training data of distributed computing nodes 130A-130N, rather than analyzing the training data itself. Thus, node analysis module 120 does not require access to the local training data itself, thereby addressing any privacy concerns regarding local training data.” Further see [0054].)
Regarding Claim 6, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
wherein the bias lineage for the particular one of the aggregated machine learning models indicates a version of at least one machine learning model on which it is based. (Thomas, [0063]-[Col. 18, Lines 15-30] “In response to a programmatic request for lineage information pertaining to one or more of the model versions… a graphical interface 601 may be used to indicate the logical relationships between the related model versions of the trials. A child or direct descendant relationship with respect to a parent model version may indicate that, for example, one or more hyper-parameters, model settings, or portions of model source code of the parent were changed to derive the child version. For example, model version 610B is shown as being derived or descended from model version 610; model versions 610C and 610D are both shown as direct descendants or children of model version 610B; and model versions 610E, 610F and 610G are shown as children of model version 610D.” [0072] “The EMS may provide lineage information about different versions of models or other experiment-related artifacts—e.g., indicating that model version V1 was an ancestor or antecedent from which model versions V2-V5 were derived, the key differences in hyper-parameters and other factors between the different versions, and so on.” Further see [Col. 5, Lines 55-65; Col. 6, Lines 1-5].) [Examiner Note: As noted above, Thomas defines the lineage graph that is generated, based on artifact/config tracking. The graph incorporates metrics as part of the lineage information presented in the same graphical interface describing the lineage records of different model versions. The “bias” lineage is defined by Das (i.e., “a display of bias metric change over time, indicating the bias metric, as well as other monitoring features.” [0122]; “Integrating the capture of information, such as feature attribution as discussed below, may provide for opportunities provide both global and local explainability for machine learning models.” [0023]).]
Regarding Original Claim 7, the combination of Watanabe, Das, and Thomas teaches the elements of claim 6 as outlined above, and further teaches:
As outlined above, Thomas, in combination with Watanabe and Das, teaches constructing bias lineages for the aggregated machine learning models based on collected metrics (e.g., bias metrics in the local training data as described by Das). The combination of Watanabe, Das, and Thomas further teaches:
further indicates which of the plurality of training nodes associated with the at least one machine learning model on which the particular one of the aggregated machine learning models is based. (Watanabe, [0051]-[0052] “Statistical data is received from each of the computing nodes at operation 220. Each computing node of distributed computing nodes 130A-130N may compute statistical data based on its local training data; the statistical data may merely describe the number of data samples for each available category of label, and as such, may be described as metadata. The statistical data received by central training server 105 may indicate a source node (e.g., an identifier of the particular node of distributed computing nodes 130A-130N from which the statistical data is obtained), and a count of data samples per label category… At operation 230, the statistical data is analyzed to identify any computing nodes whose local training data is underrepresented with respect to one or more label categories.” [0029] “Federated learning module 115 may combine the results from multiple computing nodes to generate a global trained model, which can be generated using any conventional or other federated learning techniques. In general, a global model may be generated based on parameters of the trained distributed models, such as the weights and biases of a trained distributed model's neural network; values for the global model can be obtained by averaging or otherwise combining corresponding values of multiple trained distributed models.” [0030]-[0032] “Node analysis module 120 receives or obtains statistical data from computing nodes participating in a federated learning task, and analyzes the data to identify any computing nodes whose local training data is biased beyond a threshold amount. In particular, node analysis module 120 identifies bias in local training data of distributed computing nodes 130A-130N when it is determined that some types of data samples are overrepresented or underrepresented with respect to other types of samples… node analysis module 120 identifies bias in local training data sets by comparing the count of data samples for an overrepresented label to the count of data samples for other labels.”)
Watanabe does not appear to explicitly define bias “lineage” that indicates which training node. Specifically, does not define that the source of bias information is embedded or represented in a lineage record. However, Thomas, in combination with Watanabe and Das, teaches:
wherein the bias lineage further indicates which of the plurality of training nodes associated with the at least one machine learning model on which the particular one of the aggregated machine learning models is based. (Thomas, [Col. 18, Lines 5-50] “FIG. 6 illustrates an example presentation by an experiment management service of ancestor-descendant relationships among a plurality of machine learning models used for experiment trials, according to at least some embodiments. In the depicted embodiment, one or more experiments managed with the help of the EMS may comprise at least seven trials 650A—650G), in each of which a respective version of a machine learning model is utilized. In response to a programmatic request for lineage information pertaining to one or more of the model versions (e.g., similar to the GetModelLineagelnfo request discussed in the context of FIG. 4 ), a graphical interface 601 may be used to indicate the logical relationships between the related model versions of the trials. A child or direct descendant relationship with respect to a parent model version may indicate that, for example, one or more hyper-parameters, model settings, or portions of model source code of the parent were changed to derive the child version. For example, model version 610B is shown as being derived or descended from model version 610; model versions 610C and 610D are both shown as direct descendants or children of model version 610B; and model versions 610E, 610F and 610G are shown as children of model version 610D. … in addition to showing the ancestor-descendant relationships among model versions used in various trials, a number of other types of information may be displayed. For each of the model versions a respective timestamp 650 (e.g., 650A-650G) indicating for example when the model version was created or last modified may be shown. A respective “View details” interface element 690 (e.g., 690A-690G) may be presented in some embodiments, enabling an EMS client access additional artifacts associated with the model versions, such as notes or annotations, source code, etc.” [Col. ] “In some embodiments the artifacts themselves may be distributed across multiple services of one or more provider networks, and the EMS may act as a centralized clearinghouse or aggregator for the artifacts.”) [Examiner’s Note: the lineage graph associate each model version node in the ancestor-descendant graph with the trail lineage information, including trial initiator identifier or source code. The trail initiator identifier or source code correspond to indicating which of the training nodes is associated with the ML model on which the aggregated model version (generated by the EMS as an artifact aggregator).]
The same motivation that was utilized for combining Watanabe, Das, and Thomas as set forth in claim 1 is equally applicable to claim 1.
Regarding Claim 8, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
excluding a machine learning model from a particular one of the plurality of training nodes from being used to generate one of the aggregated machine learning models, based on a determination that the bias metrics from that particular training node exceed a threshold. (Watanabe, [0032] “if the ratio of counts of one or more of the other labels to the count of most-represented label does not surpass a threshold value, then the training data set may be considered to be biased.” [0043] “training module 140 does not begin training until instructed by central training server 105, which may not transmit the instruction to a particular distributed computing node to begin training until node analysis module 120 can determine that the distributed computing node's local training data is not sufficiently biased (or the bias has been mitigated) in accordance with present invention embodiments.”) Further see [0044], [0052], and [0060].
Regarding Currently amended Claim 11,
The claim recites substantially similar limitations as corresponding claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale. Claim 1 is directed to a method, and claim 11 is directed to an apparatus,.
Watanabe also discloses an apparatus, comprising: one or more network interfaces; a processor coupled to the one or more network interfaces and configured to execute one or more processes; and a memory configured to store a process that is executable by the processor... (Watanabe, [0024]-[0025] “FIG. 1 is a block diagram depicting a computing environment 100 for training a federated learning model in accordance with an embodiment of the present invention. As depicted, computing environment 100 includes a central training server 105, distributed computing nodes 130A-130N, and a network 160…. Central training server 105 includes a network interface (I/F) 106, at least one processor 107, memory 110, and a database 128. Memory 110 may include a federated learning module 115, a node analysis module 120, and a threshold selection module 125. Central training server 105 may include a laptop computer, a tablet computer, a netbook computer, a personal computer (PC), a desktop computer, a personal digital assistant (PDA), a smart phone, a thin client, or any programmable electronic device capable of executing computer readable program instructions. Network interface 106 enables components of central training server 105 to send and receive data over a network, such as network 160. In general, central training server 105 oversees federated learning tasks, analyzes distributed computing nodes to identify any nodes having biased local training data sets, and performs actions to substantially reduce or eliminate bias during training tasks.” [0041] “Distributed computing nodes 130A-130N each include a network interface (I/F) 131, at least one processor 132, memory 135, and a database 155. Memory 135 may include a training module 140, a training data analysis module 145, and a data augmentation module 150. Distributed computing nodes 130A-130N may each include a laptop computer, a tablet computer, a netbook computer, a personal computer (PC), a desktop computer, a personal digital assistant (PDA), a smart phone, a thin client, or any programmable electronic device capable of executing computer readable program instructions.”)
Regarding Original Claim 12,
The claim recites substantially similar limitations as corresponding claim 2 and is rejected for similar reasons as claim 2 using similar teachings and rationale.
Regarding Original Claim 13,
The claim recites substantially similar limitations as corresponding claim 3 and is rejected for similar reasons as claim 3 using similar teachings and rationale.
Regarding Original Claim 15,
The claim recites substantially similar limitations as corresponding claim 5 and is rejected for similar reasons as claim 5 using similar teachings and rationale.
Regarding Original Claim 16,
The claim recites substantially similar limitations as corresponding claim 6 and is rejected for similar reasons as claim 6 using similar teachings and rationale.
Regarding Original Claim 17,
The claim recites substantially similar limitations as corresponding claim 7 and is rejected for similar reasons as claim 7 using similar teachings and rationale.
Regarding Original Claim 18,
The claim recites substantially similar limitations as corresponding claim 8 and is rejected for similar reasons as claim 8 using similar teachings and rationale.
Regarding Currently amended Claim 20,
The claim recites substantially similar limitations as corresponding claim 1 and is rejected for similar reasons as claim 1 using similar teachings and rationale. Claim 1 is directed to a method, and claim 20 is directed to a non-transitory, computer-readable medium.
Watanabe also discloses a non-transitory, computer-readable medium storing program instructions that cause a device to execute a process... (Watanabe, [0095]-[0096] “The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention. The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device.”)
Claim(s) 4 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Watanabe, Das, and Thomas as outlined above, and further in view of Bhide et al., (Pub. No.: US 20210357803 A1).
Regarding Original Claim 4, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
As outlined above, Thomas, in combination with Watanabe and Das, teaches constructing bias lineages for the aggregated machine learning models based on collected metrics (e.g., bias metrics in the local training data as described by Das). The combination of Watanabe, Das, and Thomas further teaches:
wherein the bias … for the particular one of the aggregated machine learning models indicates at least one of the local training datasets as a source of bias for a data feature used by that aggregated machine learning model. (Watanabe, [0051]-[0052] “Statistical data is received from each of the computing nodes at operation 220. Each computing node of distributed computing nodes 130A-130N may compute statistical data based on its local training data; the statistical data may merely describe the number of data samples for each available category of label, and as such, may be described as metadata. The statistical data received by central training server 105 may indicate a source node (e.g., an identifier of the particular node of distributed computing nodes 130A-130N from which the statistical data is obtained), and a count of data samples per label category… At operation 230, the statistical data is analyzed to identify any computing nodes whose local training data is underrepresented with respect to one or more label categories.” [0019] “Accordingly, present invention embodiments identify any computing nodes participating in a federated learning task whose local training data sets may be biased to some degree, and mitigate or eliminate bias in order to improve the resulting global model.” [0030]-[0032] “Node analysis module 120 receives or obtains statistical data from computing nodes participating in a federated learning task, and analyzes the data to identify any computing nodes whose local training data is biased beyond a threshold amount. In particular, node analysis module 120 identifies bias in local training data of distributed computing nodes 130A-130N when it is determined that some types of data samples are overrepresented or underrepresented with respect to other types of samples… node analysis module 120 identifies bias in local training data sets by comparing the count of data samples for an overrepresented label to the count of data samples for other labels.”)
Watanabe does not appear to explicitly define “lineage.” Specifically, does not define that the source of bias information is embedded or represented in a lineage record. However, Bhide, in combination with Watanabe, Das, and Thomas, teaches:
wherein the bias lineage for the particular one of the aggregated machine learning models indicates at least one of the local training datasets as a source of bias for a data feature used by that aggregated machine learning model. (Bhide, [0038] “the lineage of the predictive model training data 110 used to build the predictive model 114 is captured 306 during the building and training of the predictive model 114, and is identified as the training data lineage metadata 122. The training data lineage metadata 122 includes, without limitation, any dates associated with the respective data, any labels associated with the respective data established through the model training, column, row, and table identifiers, original source of the data, and any respective linking or joining features.” [0024] “predictive model design time measurements are captured during the model creation and training activities. Such design time measurements include metrics of the data used to construct and train the model such as, without limitation, accuracy, precision, recall, bias, and drift. In the respective embodiments, these design time metrics are correlated to the respective features which were culled from the respective data to build the model. Accordingly, data metrics are used to assign quality metrics to the features.” [0045] “one or more runtime analyses directed at identifying 410 those features of the features 226 associated with the predictive model 214 that tend to introduce indirect biases in the predictive model 214 are executed. In one or more embodiments, it may be determined that a particular feature being evaluated is not being used in the predictive model 214, but is correlated to some other feature or features which are used in the predictive model 214 and tend to direct the predictive model 214 to perform in a biased manner. Accordingly, features 226 that introduce biases in the predictive model 214 may be removed, or, alternatively, the predictive model 214 may be calibrated to substantially negate such biases.”)
Accordingly, at the effective filing date, it would have been prima facie obvious to one ordinarily skilled in the art to modify the combination of Watanabe, Das, and Thomas to incorporate the method for generating an enhanced feature catalog for a predictive model as taught by Bhide. One would have been motivated to make such a combination in order to improve the known art of building and validating machine learning models through automatically determining the relevant features for generating predictions and decisions through the models, and significantly reduces the time and effort necessary to determine such relevant features (Bhide [0055]).
Regarding Original Claim 14,
The claim recites substantially similar limitations as corresponding claim 4 and is rejected for similar reasons as claim 4 using similar teachings and rationale.
Claim(s) 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over the combination of Watanabe, Das, and Thomas as outlined above, and further in view of Bhide et al., (Pub. No.: US 20220101180 A1).
Regarding Original Claim 9,
the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
As outlined above, Watanabe teaches the concept of excluding a machine learning model from a particular training node from being used to generate the aggregated global model based on a bias-threshold determination (see Watanabe [0032] and [0043]). Das and Thomas teaches the lineage tracking of model versions with its corresponding metadata using lineage graph and the display of the model bias metrics change over time. The combination of Watanabe, Das, and Thomas does not appear to explicitly suggest that the lineage indicates that the machine learning model was excluded.
However, it would have been obvious in view of Bhide. Hereinafter, Bhide, in combination with Watanabe, Das, and Thomas, teaches:
wherein the bias lineage indicates that the machine learning model from the particular one of the plurality of training nodes was excluded from being used to generate one of the aggregated machine learning models. (Bhide, [0016]-[0017] “To keep track of these various events, a lineage path may be used. A lineage path, as referred to herein, may be a recordation (e.g., storing) of the various lineage events of the machine learning model. … the lineage path may be a lineage tree (e.g., similar to a decision tree).” [0090] “In some instances, after a new version of the ML model v1 424 was created, the ML model v1 may eventually be demoted, or removed, from the production environment. This is represented as v1 model demotion 431.” Further see [0020] and [0093].)
Accordingly, it would have been obvious to a person having ordinary skill in the art, before the effective filing date of the claimed invention, having the combination of Watanabe, Das, Thomas, and Bhide, to incorporate the method for generating lineage events of machine learning models as taught by Bhide. One would have been motivated to make such a combination in order to help improve (e.g., increase efficiency, accuracy, etc.) the model throughout its lifecycle. Additionally, keeping track of the lineage events may help ensure credibility and transparency of the machine learning model when the model is deployed in production (Bhide [0019]).
Regarding Original Claim 19,
The claim recites substantially similar limitations as corresponding claim 9 and is rejected for similar reasons as claim 9 using similar teachings and rationale.
Claim(s) 10 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Watanabe, Das, and Thomas as outlined above, and further in view of Xing et al., (NPL: "Jupiter: a modern federated learning platform for regional medical care." (2021)).
Regarding Original Claim 10, the combination of Watanabe, Das, and Thomas teaches the elements of claim 1 as outlined above, and further teaches:
As outlined above, Watanabe teaches generating aggregated machine learning models (global models) over time via iterative federated learning (see [0029] and [0037]). Das and Thomas teaches the lineage tracking of model versions with its corresponding metadata using lineage graph and the display of the model bias metrics change over time. The combination of Watanabe, Das, and Thomas does not appear to explicitly suggest rolling back the aggregated model to a prior version in response to a user interface request.
However, it would have been obvious in view of Xing. Hereinafter, Xing, in combination with Watanabe, Das, and Thomas, teaches:
rolling back the particular one of the aggregated machine learning models to a prior version, in response to a request to do so from a user interface, based in part on the bias lineage for the particular one of the aggregated machine learning models. (Xing, [P. 5, Section: 3.2] “Developers create their federated learning job (FL job) in the IDE using specifications with respect to model structure, data provisioning and hyperparameters/metrics, submit this job to the controller and finally inspect feedback for metrics.” [P. 7, Section: 3.4] “Manipulation APIs are intended for the incremental tunning of processing pipelines and datasets; the rollback API provides the ability to convert FL Data into a specific history checkpoint; tracking APIs are used to shape the lineage based on other FL Data lineages.” [P. 8, Section: 3.4, Table 1] “Manipulation APIs: … X.rollback(history number)… Tracking APIs: X.lineage(indexes), X.diff(Y), X.merge(Y), X.freeze(indexes)(Y).” [P. 8, Section: 3.4] “A session typically maintains the following states into a committed log history … Subsequent tuning activities incrementally change the state space and forms a log history.” [P. 7, Section: 3.2] “Once the developer tunes the FL job (e.g., rerendering datasets, tunning hyperparameters), she can modify the corresponding specifications and update the job with the session token. Specifically FL Data abstraction (Subsection 3.4).”) [Examiner’s Note: Rollback API imitated via the developer IDE that rolls back the federated learning job/model state to a specific prior history checkpoint from committed log history/lineage.]
Therefore, it would have been prima facie obvious to one of ordinary skill in the art, before the effective date of the claimed invention, having the combination of Watanabe, Das, Thomas, and Xing to incorporate the proposed Jupiter framework API functions as taught by Xing. One would have been motivated to make such a combination in order to construct a secure and high performance aggregator cluster with SGX to efficiently aggregate the encrypted model parameters. Thereby providing an easy-to-use, secure, and high-performance platform for federated machine learning (Xing [Abstract]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
(Pub. No.: US 20220012591 A1) – “Angelo Dalli” relates to “Method for detecting and mitigating bias and weakness in artificial intelligence training data and models.”
[Abstract] “Methods for detecting bias, strength, and weakness of data sets and the resulting models may be described. A method may implement global bias detection which utilizes the coefficients of the model to identify, minimize, and/or correct potential bias within a desired error tolerance. Another method makes use of local feature importance extracted from the rule-based model coefficients to locally identify bias. A third method aggregates the feature importance over the results/explanations of multiple samples. Bias may also be detected in multi-dimensional data such as images. A backmap reverse indexing mechanism may be implemented. A number of mitigation methods are also presented to eliminate bias from the affected models.”
(Pub. No.: US 20220172004 A1) – “Sanjiv Das” relates to “Monitoring bias metrics and feature attribution for trained machine learning models.”
[Abstract] “A data lineage system tracks performance of data flows through different transformations independent of the systems that perform the transformations. A data flow model is maintained as a graph in the data lineage system that is updated by data processors to include performance history of different transformations in the data flow. Subsequent analyses of the data flow model, such as tracing particular data, can be supported using the recorded performance information in the graph of the data flow model.”
(Pub. No.: US 11755536 B1) – “Yahor Pushkin” relates to “System-independent data lineage system.”
[Abstract] “A data lineage system tracks performance of data flows through different transformations independent of the systems that perform the transformations. A data flow model is maintained as a graph in the data lineage system that is updated by data processors to include performance history of different transformations in the data flow. Subsequent analyses of the data flow model, such as tracing particular data, can be supported using the recorded performance information in the graph of the data flow model.”
NPL: Biswas, Debmalya, and Krishnamurthy Vidyasankar. "A Privacy Framework for Hierarchical Federated Learning." (2021).
[Abstract]: “In this paper, we propose a conceptual framework for HFL by capturing the data/model at tributes at each node, including their privacy exposure. The framework enables scenarios where a node output may expose certain at tributes of its underlying data, as well as identifying models in the hierarchy that need to be updated once a user whose data was used in their training has opted-out.”
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SADIK ALSHAHARI whose telephone number is (703)756-4749. The examiner can normally be reached Monday - Friday, 9 a.m. 6 p.m. ET.
Examiner interviews are available via telephone, 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, Li Zhen can be reached on (571) 272-3768. 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.
/S.A.A./Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
