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 .
Claims 1-10 are presented for examination.
Specification
The disclosure is objected to because of the following informalities:
Page 3 Lines 16-18: the limitation “A model can longer memorize a class having high accuracy...” should read "A model can better memorize" for clarity.
Page 7 Lines 16-17: the limitation "a classification mode can easily perform classification by reducing a class dispersion" should read "a classification model can easily perform classification…".
Page 16 Lines 5-6: The cited reference is recited as "iCarl". It should read as "iCaRL" for consistency with the reference title.
Appropriate correction is required.
The use of the term BLUETOOTH, which is a trade name or a mark used in commerce, has been noted in this application. The term should be accompanied by the generic terminology; furthermore, the term should be capitalized wherever it appears or, where appropriate, include a proper symbol indicating use in commerce such as ™, SM , or ® following the term.
Although the use of trade names and marks used in commerce (i.e., trademarks, service marks, certification marks, and collective marks) is permissible in patent applications, the proprietary nature of the marks should be respected and every effort made to prevent their use in any manner which might adversely affect their validity as commercial marks.
Claim Objections
Claim 6 and its respective dependent claims are objected to because of the following informality:
Claim 6 recites "A computer device for fair few-shot shot class-incremental learning (CIL)," which should instead read "A computer device for fair few-shot class-incremental learning (CIL).".
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-10 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
The term “fair” in claims 1, 6, and 10 is a relative term which renders the claim indefinite. The term “fair” is not defined by the claim, the specification does not provide a standard for ascertaining the requisite degree, and one of ordinary skill in the art would not be reasonably apprised of the scope of the invention. Although the specification discusses fairness in terms of an accuracy difference between super classes (overall accuracy equality) and a "fairness target," it does not set forth any standard for determining the degree to which the claimed learning must be "fair," such that the metes and bounds of the term cannot be ascertained. Claims 2-5 and 7-9 depend from claims 1 and 6 and are rejected for the same reason.
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 1-10 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an
abstract idea without significantly more.
Claim 1
Step 1: The claim recites a method; therefore, it is directed to the statutory category of a process.
Step2A Prong 1: The claim recites, inter alia:
constructing a separate storage … by using samples of a first super class having first accuracy and samples of a second super class having second accuracy lower than the first accuracy: This limitation recites a mental process because selecting and grouping samples of a first super class and a second super class based on an evaluation of their respective accuracies can be performed in the human mind or by a human using pen and paper through evaluation, judgement, and opinion.
adjusting a number of samples of the first super class and a number of samples of the second super class in the separate storage … when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning: This limitation recites a mental process because determining and adjusting the number of samples of each super class based on whether the results satisfy a fairness criterion or the comparison of the first accuracy to the second accuracy is an evaluation and judgement that can be performed in the human mind or by a human using pen and paper.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
constructing a separate storage device: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself (MPEP 2106.05(h)).
performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
constructing a separate storage device: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself which cannot provide inventive concept (MPEP 2106.05(h)).
fair few-shot class-incremental learning (CIL) of a computer device, the method comprising… performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea and cannot provide inventive concept (MPEP 2106.05(f)).
The elements in combination as an ordered whole still do not amount to significantly more than the judicial exception (i.e., the abstract ideas of mental processes for selecting and grouping training samples of a first super class and a second super class according to their respective accuracies and adjusting the numbers of samples based on evaluating a fairness criterion and comparing accuracies). The claim merely describes a process of applying known organizational and evaluative techniques to construct an exemplar set and repeatedly train on it, along with a generic storage device and generic incremental learning that merely indicate a technological environment in which the abstract ideas are applied without improving the functioning of a computer or the incremental learning model itself.
Therefore, the claim as a whole remains focused on the abstract idea and fails Step 2B of the eligibility analysis.
Claim 2
Step 1: A process, as above.
Step2A Prong 1: The claim recites, inter alia:
the constructing of the separate storage … comprises storing the samples of the second super class by a maximum number that is permitted with respect to the second super class in the separate storage …: This limitation recites a mental process because it involves capping the number of samples stored in a group, which can be performed in the human mind or by pen and paper.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
the constructing of the separate storage device comprises … storing the samples of the second super class … in the separate storage device: This limitation is merely a post-solution step of storing the data—a nominal addition to the claim that does not meaningfully limit the claim. The method storing is recited at a high level of generality. Simply implementing the abstract idea in a generic method is not a practical application of the abstract idea. Therefore, storing step is an insignificant extra-solution activity. See MPEP 2106.05(g).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
the constructing of the separate storage device comprises … storing the samples of the second super class … in the separate storage device: These elements amount to storing… information in memory, Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93; See MPEP 2106.05(d) (II)(iv). The courts have recognized the computer functions of storing as well‐understood, routine, and conventional function when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity.
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim 3
Step 1: A process, as above.
Step2A Prong 1: The claim recites, inter alia:
returning to adjusting the number of samples of the first super class and the number of samples of the second super class…: This limitation recites a mental process because returning to and repeating the evaluation and judgement of adjusting the numbers of samples of the first and second super classes can be performed in the human mind or by pen and paper.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
after performing the incremental learning in the next step: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
after performing the incremental learning in the next step: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea and cannot provide inventive concept (MPEP 2106.05(f)).
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim 4
Step 1: A process, as above.
Step2A Prong 1: The claim recites, inter alia:
the adjusting of the number of samples of the first super class and the number of samples of the second super class and performing the incremental learning in the next step are repeated when the results of the incremental learning do not satisfy the fairness criterion and the first accuracy is higher than the second accuracy in the results of the incremental learning: This limitation recites a mental process because it involves deciding when to repeat the adjusting of the numbers of samples based on whether the results fail to satisfy the fairness criterion and the accuracy comparison.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
and performing the incremental learning in the next step are repeated: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
and performing the incremental learning in the next step are repeated: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea and cannot provide inventive concept (MPEP 2106.05(f)).
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim 5
Step 1: A process, as above.
Step2A Prong 1: The claim recites, inter alia:
selecting a separate storage device having the number of samples of the first super class and the number of samples of the second super class when overall accuracy is a highest in the results of the incremental learning: This limitation recites a mental process because it involves the evaluation/judgement/opinion of selecting the configuration of samples for which the overall accuracy is the highest.
Step 2A Prong Two and Step 2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception under step 2B. Thus, the judicial exception is not integrated into a practical application (see MPEP 2106.04(d) I.), failing step 2A prong 2. The claim is ineligible.
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim 6
Step 1: The claim recites a computer device; therefore, it is directed to the statutory category of a machine.
Step2A Prong 1: The claim recites, inter alia:
construct a separate storage … by using samples of a first super class having first accuracy and samples of a second super class having second accuracy lower than the first accuracy: This limitation recites a mental process because selecting and grouping samples of a first super class and a second super class based on an evaluation of their respective accuracies can be performed in the human mind or by a human using pen and paper through evaluation, judgement, and opinion.
adjust a number of samples of the first super class and a number of samples of the second super class in the separate storage … when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning: This limitation recites a mental process because determining and adjusting the number of samples of each super class based on whether the results satisfy a fairness criterion or the comparison of the first accuracy to the second accuracy is an evaluation and judgement that can be performed in the human mind or by a human using pen and paper.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
[a] computer device for fair few-shot shot class-incremental learning (CIL), comprising: memory; and a processor connected to the memory and configured to execute at least one instruction stored in the memory, wherein the processor is configured to … performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).
construct a separate storage device: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself (MPEP 2106.05(h)).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
[a] computer device for fair few-shot shot class-incremental learning (CIL), comprising: memory; and a processor connected to the memory and configured to execute at least one instruction stored in the memory, wherein the processor is configured to … performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea and cannot provide inventive concept (MPEP 2106.05(f)).
construct a separate storage device..: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself which cannot provide inventive concept (MPEP 2106.05(h)).
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim 7 recites similar limitations to claim 2. Therefore, claim 7 is rejected using the same rationale as claim 2.
Claim 8 recites similar limitations to claims 3 and 4. Therefore, claim 8 is rejected using the same rationale as claims 3 and 4.
Claim 9 recites similar limitations to claim 5. Therefore, claim 9 is rejected using the same rationale as claim 5.
Claim 10
Step 1: The claim recites a non-transitory computer-readable recording medium; therefore, it is directed to the statutory category of an article of manufacture.
Step2A Prong 1: The claim recites, inter alia:
constructing a separate storage … by using samples of a first super class having first accuracy and samples of a second super class having second accuracy lower than the first accuracy: This limitation recites a mental process because selecting and grouping samples of a first super class and a second super class based on an evaluation of their respective accuracies can be performed in the human mind or by a human using pen and paper through evaluation, judgement, and opinion.
adjusting a number of samples of the first super class and a number of samples of the second super class in the separate storage … when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning: This limitation recites a mental process because determining and adjusting the number of samples of each super class based on whether the results satisfy a fairness criterion or the comparison of the first accuracy to the second accuracy is an evaluation and judgement that can be performed in the human mind or by a human using pen and paper.
Step2A Prong 2: This judicial exception is not integrated into a practical application because the
additional elements are as follows:
A non-transitory computer-readable recording medium storing at least one program to execute a method for fair few-shot class-incremental learning (CIL) in a computer device, wherein the method for fair few-shot CIL comprises… performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f)).
constructing a separate storage device: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself (MPEP 2106.05(h)).
Step 2B: The claim does not include additional elements that are sufficient to amount to significantly
more than the judicial exception because the additional elements are as follows:
A non-transitory computer-readable recording medium storing at least one program to execute a method for fair few-shot class-incremental learning (CIL) in a computer device, wherein the method for fair few-shot CIL comprises… performing incremental learning on the separate storage device… performing incremental learning in a next step on the separate storage device: Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea and cannot provide inventive concept (MPEP 2106.05(f)).
constructing a separate storage device: The limitation amounts to merely indicating a field of use or technological environment in which to apply a judicial exception. This does not amount to significantly more than the exception itself which cannot provide inventive concept (MPEP 2106.05(h)).
Even when considered in combination, these additional elements represent mere instructions to apply
an exception and therefore do not provide an inventive concept. The claim is ineligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1, 3-6, and 8-10 are rejected under 35 U.S.C. 103 as being unpatentable over Liu (“RMM: Reinforced Memory Management for Class-Incremental Learning”, January 2023), in view of Chowdhury (“Sustaining Fairness via Incremental Learning”, 2022), in view of Kumar (US 20240362419 A1), and in further view of Goto (US 11934944 B2).
Regarding claim 1,
Liu teaches constructing a separate storage … by using samples of a first super class having first [entropy] and samples of a second super class having second [entropy] lower than the first [entropy] (Page 2 Introduction, “Level-1 function determines how to split memory between the old and new data. Its output action is then inputted into the Level-2 function to determine how to allocate memory for each old class. The overall objective of the function is to maximize the cumulative evaluation accuracy across all incremental phases.”, Page 5 Section 4.1 of Liu, “we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group. Therefore, Level-2 action a[2]i ∈ (0,1) determines how to split memories between harder and easier classes.”
Liu teaches splitting the phase's classes into two groups by training entropy, the higher-entropy difficult classes in one group and the remaining easier classes in the other. These groups are saved into its exemplar-memory partition. The two entropy groups correspond to the two different super classes. The second super class corresponds to Liu's high-entropy difficult group because Liu treats higher entropy as a more difficult class.)
performing incremental learning on the separate storage … (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew. We select Et as representative samples of the data seen in the t-th phase… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di. Then, we initialize Θi with Θi−1, and train it using E0:i−1 ∪ Di. The resulting model Θi will be evaluated with a test set containing all classes observed so far. We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches that in each phase it initializes the current model from the prior phase's model and trains on the union of the stored exemplars and the new data, i.e., replaying the exemplars held in its exemplar memory. Training on those stored exemplars corresponds directly to performing incremental learning on the separate storage.);
and performing incremental learning in a next step on the separate storage … (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches iterating across all incremental phases, repeating the training on the updated exemplar memory in each successive phase. Re-running the replay training in the following phase corresponds to performing incremental learning in a next step on the separate storage.).
adjusting a number of samples of the first super class and a number of samples of the second super class in the separate storage … (Page 3 Section 3, “an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… M is used to store the exemplars and new coming data as both kinds of data need to be loaded repeatedly during training epochs. In the initial (0-th) phase, data D0, containing the training samples of C0 classes, are used to learn the initial classification model Θ0. In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di… We repeat this training and testing, and report the average accuracy across all phases.”, Page 5 Level-2 Actions, “In the (i−1)-th phase, we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group… Let MA j and MB j denote the memory allocated for the high-entropy and low-entropy groups, respectively, in the j-th phase (j ≤ i): [Equation 2]. Then, we allocate memory evenly to the classes within the group, e.g., if the high-entropy group has 10 classes, each class will have a memory size of 1/10 |M_j|.”, See Figure 1(b) and its Caption,
PNG
media_image1.png
342
471
media_image1.png
Greyscale
”Our proposed method— Reinforced Memory Management (RMM)—is able to learn the optimal and class-specific memory sizes in different incremental phases.”, Figure 2(b) caption, “For the k-th pseudo CIL task, we allocate memory for N times (i.e., in N phases) using the policies pi_n and pi_phi, and compute the cumulative reward R.”
Liu teaches that in each successive phase the Level-2 action re-allocates the memory between the harder and easier groups after the model has been trained in the prior phase, and the algorithm loops back to re-allocate for the next phase.)
Liu does not teach [a] method for fair few-shot class-incremental learning (CIL) of a computer device, the method comprising: constructing a separate storage device… performing incremental learning on the separate storage device… when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning;
Chowdhury, in the same field of endeavor, teaches [performing an action] when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning (Page 1 Introduction, “To address this problem, we propose a representation learning system– Fairness-aware Incremental Representation Learning (FaIRL). At its core, FaIRL uses an adversarial debiasing setup for removing demographic information by controlling the number of bits (rate-distortion) required to encode the learned representations… We leverage this debiasing setup for incremental learning using an exemplar-based approach, by retaining a small set of representative samples from previous tasks, to prevent forgetting... We propose FaIRL, a representation learning system that learns fair representations, while incrementally learning new tasks, by controlling their rate-distortion function.”, Page 4 Incremental Learning, “To ensure fairness, the system also needs to learn representations that are oblivious to the protected attribute g for both X_new and X_old… This is achieved by minimizing the discriminator loss ∆R(Znew, Πg new) … The system should learn fair representations for Xold. This is achieved by minimizing the discriminator loss for the exemplars ∆R (Zold, Πg old) (Equation4). The overall objective function that the encoder optimizes in the incremental learning setup: [Equation 6]”, Page 5 Metrics, “TPR-GAP (De-Arteaga et al. 2019) computes the difference between true positive rates between two protected groups Gap (g, y) = TPR_(g, y) – TPR_(g, y), where g, ¯g are possible values of the protected attribute.”
Chowdhury teaches an exemplar-based incremental learning system whose objective is fairness across two protected groups, quantified by the true positive rate gap between those groups. In response to that fairness measure, the system is updated at each training stage by minimizing the discriminator loss for both the new data and the stored exemplars. Evaluating the fairness measure and then performing the responsive update corresponds to performing an action when the results of the incremental learning are assessed against the fairness criterion.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu's memory allocation method with Chowdhury's fairness objective in order to determine when to adjust the per super class sample counts to reduce the accuracy disparity between the two groups (Page 4 Incremental Learning of Chowdhury).
Liu in view of Chowdhury does not teach [a] method for fair few-shot class-incremental learning (CIL) of a computer device, the method comprising: constructing a separate storage device… performing incremental learning on the separate storage device.
Kumar, in the same field of endeavor, teaches [a] method for fair few-shot class-incremental learning (CIL) of a computer device, the method comprising: constructing a separate storage device… performing incremental learning on the separate storage device (Paragraph 12, “…few-shot incremental learning may be achieved to train a model which accurately classifies for both old and new classes. Catastrophic forgetting of the original classes may be avoided while the addition of new classes is supported. Few-shot incremental learning may be achieved with a smaller training dataset for a new class so that time and resources are saved with respect to annotation/labeling requirements.”, Paragraph 37, “A computer program product … is a term … to describe any set of one, or more, storage media (also called “mediums”) collectively included in a set of one, or more, storage devices that collectively include machine readable code corresponding to instructions and/or data for performing computer operations”, Paragraph 70, “In step 306 of the few-shot incremental deep learning model training process 300 shown in FIG. 3, last hidden state features corresponding to tokens from first and second classes are saved…For example, for a first prototype of a first class, a first set of last hidden state features may be saved and stored in a database… For a second prototype of the second class, another different set, e.g., a fourth set, of last hidden state features may be saved and stored in the database… the database may be within the persistent storage 113 of the client computer”, See Figure 1.
Kumar teaches a few-shot incremental learning that classifies both old and new classes while avoiding catastrophic forgetting using a small new-class dataset, implemented on a computer program product with storage devices. The prototype state features are stored in a database in the computer's storage.)
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury's class-incremental learning with Kumar's few-shot incremental learning implemented on a computer having storage in order to add new classes into a storage device while saving annotation time and resources (Paragraph 12 of Kumar).
Liu in view of Chowdhury and in further view of Kumar does not teach a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy.
Goto, in the same field of endeavor, teaches [identifying] a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy (Col. 6 Lines 2-7 of Goto, “For example, if the accuracy of class A is 0.95 while the accuracy of class B is 0.1, the magnitude of augmentation for class B (L.sub.B) may be updated (e.g., L.sub.B=L.sub.B+1) so that the number of training samples of class B, L.sub.BM.sub.B is increased in the next training loop as compared to class A, L.sub.AM.sub.A.”
Goto teaches evaluating class accuracy after training and identifying two classes as a first class having greater accuracy and a second class having less accuracy, expressly comparing the two, e.g., a class A with an accuracy of 0.95 against a class B with an accuracy of 0.1. Goto further teaches increasing the number of training samples of the lower-accuracy second class relative to the higher-accuracy first class in the next training loop.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury in further view of Kumar’s incremental per-group memory allocation with Goto's accuracy difference driven update condition in order to characterize Liu's two exemplar-memory groups by their measured class accuracy (Col. 6 Lines 2-12 of Goto).
Regarding claim 3,
Liu teaches returning to adjusting the number of samples of the first super class and the number of samples of the second super class after performing the incremental learning in the next step (Page 7 Algorithm 1 Lines 9-17, “for i in 0,...,N do… if i ≥ 1 do… Observe si and produce a[1] ∼πη (s_i)… Allocate M_old and M_new using Eq. 1… Produce a[2] ∼πφ( a[1]_i , s_i)… Allocate {MA j}i j=0 and {MB j}_i j = 0 using Eq. 2… Update E0: i−1 using herding [40]… Save E0: i−1 in M_old and free M_new; Observe new data and load Di into M_new randomly;”, See Figure 1(b) and its Caption,
PNG
media_image1.png
342
471
media_image1.png
Greyscale
”Our proposed method— Reinforced Memory Management (RMM)—is able to learn the optimal and class-specific memory sizes in different incremental phases.”, Figure 2(b) caption, “For the k-th pseudo CIL task, we allocate memory for N times (i.e., in N phases) using the policies pi_n and pi_phi, and compute the cumulative reward R.”
Liu teaches that in each successive phase the Level-2 action re-allocates the memory between the harder and easier groups after the model has been trained in the prior phase, and the algorithm loops back to re-allocate for the next phase. Re-running the group-memory allocation in each subsequent phase corresponds to returning to adjusting the number of samples of the first and second super class after performing the incremental learning in the next step.).
Regarding claim 4,
Liu teaches the adjusting of the number of samples of the first super class and the number of samples of the second super class and performing the incremental learning in the next step are repeated… (Page 3 Section 3, “In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di… We repeat this training and testing, and report the average accuracy across all phases.”, See Figure 1(b) and its Caption,
PNG
media_image1.png
342
471
media_image1.png
Greyscale
”Our proposed method— Reinforced Memory Management (RMM)—is able to learn the optimal and class-specific memory sizes in different incremental phases.”
Liu teaches that in each successive phase the Level-2 action re-allocates the memory between the harder and easier groups after the model has been trained in the prior phase, and the algorithm loops back to re-allocate for the next phase.)
Liu does not teach …when the results of the incremental learning do not satisfy the fairness criterion and the first accuracy is higher than the second accuracy in the results of the incremental learning.
Chowdhury, in the same field of endeavor, teaches performing the incremental learning in the next step are repeated when the results of the incremental learning do not satisfy the fairness criterion… (Page 3 Figure 2 Caption, “The feature encoder tries to learn discriminative representations for the target task (y) using MCR2 objective while minimizing the discriminator loss.”, Page 3 Fairness-aware Incremental Representation learning (FaIRL), “In FaIRL, we empirically observe that the feature spaces learned at each training stage are compact. This happens because while learning discriminative representations using the MCR2 objective (∆R (Z, Πy))… Minimizing ∆R (Z, Πg) makes representations from different protected classes similar, resulting in a compact feature space. The ∆R (Z, Πg) term acts as a natural regularizer to the MCR2 objective, and prevents the learned representations from expanding in an un constrained manner, making them suitable for incremental learning.”, Page 3-4 Incremental Learning, “For incremental learning, we use an exemplar-based approach… We store a small set of exemplars from old tasks Xold = {X1 old, ..., X_old}, where m is the number of target classes (m = c (t−1)) … the system has encountered so far (each training step introduces c target classes, k is the total number of classes). The goal of our system is to learn discriminative representations w.r.t y for Xnew while retaining the old representation subspaces of Xold. To ensure fairness, the system also needs to learn representations that are oblivious to the protected attribute g for both Xnew and Xold… The system should learn fair representations for Xold. This is achieved by minimizing the discriminator loss for the exemplars ∆R (Zold, Πg old) (Equation4).”, Page 5 Metrics, “TPR-GAP (De-Arteaga et al. 2019) computes the difference between true positive rates between two protected groups Gap (g, y) = TPR (g, y) – TPR_(g, y), where g, ¯g are possible values of the protected attribute.”
Chowdhury teaches an exemplar-based incremental learning system whose objective is fairness across two protected groups, quantified by the true-positive-rate gap between those groups. A gap that remains above the fair state corresponds to the results not satisfying the fairness criterion, such that the system continues to reduce the disparity across training stages until fairness is achieved (i.e. minimizing the discriminator loss).)
Liu in view of Chowdhury in further view of Kumar does not teach performing the [machine] learning in the next step are repeated when … the first accuracy is higher than the second accuracy in the results of the [machine] learning.
Goto, in the same field of endeavor, teaches performing the … learning in the next step are repeated when … the first accuracy is higher than the second accuracy in the results of the … learning (Col. 14 Lines 22-50 Claim 7 of Goto, “A system for training a neural network with augmented data… obtain a dataset for a plurality of classes for training the neural network… wherein an amount of augmented data generated is determined by a data augmentation variable; train the neural network for the plurality of classes with the augmented dataset produced according to the data augmentation variable… determine a total loss and a difference of class accuracy for each class based on results of training the neural network, wherein the difference of class accuracy is a difference between the determined accuracy of different classes; update the data augmentation variable based on the total loss and class accuracy for each class… wherein augmenting the dataset according to the updated data augmentation variable includes: comparing the difference of class accuracy for a first class with greater accuracy and a second class with less accuracy to a threshold…and train the neural network with the augmented dataset produced according to the updated data augmentation variable, wherein operations of updating, augmenting according to the updated data augmentation variable, and training with the augmented dataset produced according to the updated data augmentation variable are performed until a ratio of a first number of augmented samples over a second number of augmented samples is equal to or larger than a ratio threshold…”
Goto teaches comparing the accuracy of a first higher-accuracy class against a second lower-accuracy class. When the first accuracy exceeds the second by more than a threshold, the model increases the number of samples of the lower-accuracy class in the next training iteration. Goto further teaches repeating this update and retrain loop while the higher class exceeds the lower class and ceasing once the accuracy difference falls below the threshold. This directly maps to repeating the adjusting and the learning in the next step when the first accuracy is higher than the second accuracy.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury in further view of Kumar’s incremental per-group memory allocation with Goto's accuracy difference driven update condition in order to iteratively reduce the accuracy disparity between the two groups until they are balanced (Col. 14 Lines 22-50 of Goto).
Regarding claim 5,
Liu teaches selecting a separate storage … having the number of samples of the first super class and the number of samples of the second super class when overall accuracy is a highest in the results of the incremental learning (Page 5 Rewards, “In the i-th phase, the objective of RMM is to maximize the expected cumulative reward, i.e., R = N i=0 ri, where ri denotes the validation accuracy in the i-th phase.”, Page 6 Training, “Training. We elaborate the steps of learning Level-1 policy πη and Level-2 policy πφ in the following. The goal is to optimize the parameters η and φ by maximizing the expected cumulative reward…”, See Algorithm 1.
Liu teaches that the memory-allocation policy which sets the per group memory sizes, is optimized to maximize the cumulative validation accuracy and the allocation yielding the highest accuracy is selected. Selecting the memory allocation that produces the highest overall accuracy corresponds to selecting the separate storage device having the numbers of samples of the first and second super class when the overall accuracy is highest.).
Liu in view of Chowdhury does not teach selecting a separate storage…
Kumar, in the same field of endeavor, teaches selecting a separate storage device (Paragraph 37, “A computer program product … is a term … to describe any set of one, or more, storage media (also called “mediums”) collectively included in a set of one, or more, storage devices that collectively include machine readable code corresponding to instructions and/or data for performing computer operations”, Paragraph 70, “In step 306 of the few-shot incremental deep learning model training process 300 shown in FIG. 3, last hidden state features corresponding to tokens from first and second classes are saved…For example, for a first prototype of a first class, a first set of last hidden state features may be saved and stored in a database… the database may be within the persistent storage 113 of the client computer”)
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury's class-incremental learning with Kumar's few-shot incremental learning implemented on a computer having storage in order to add new classes into a storage device while saving annotation time and resources (Paragraph 12 of Kumar).
Regarding claim 6,
Liu teaches construct a separate storage … a separate storage … by using samples of a first super class having first [entropy] and samples of a second super class having second [entropy] lower than the first [entropy] (Page 2 Introduction, “Level-1 function determines how to split memory between the old and new data. Its output action is then inputted into the Level-2 function to determine how to allocate memory for each old class. The overall objective of the function is to maximize the cumulative evaluation accuracy across all incremental phases.”, Page 5 Section 4.1 of Liu, “we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group. Therefore, Level-2 action a[2]i ∈ (0,1) determines how to split memories between harder and easier classes.”
Liu teaches splitting the phase's classes into two groups by training entropy, the higher-entropy difficult classes in one group and the remaining easier classes in the other. These groups are saved into its exemplar-memory partition. The two entropy groups correspond to the two different super classes. The second super class corresponds to Liu's high-entropy difficult group because Liu treats higher entropy as a more difficult class.),
perform incremental learning on the separate storage … (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew. We select Et as representative samples of the data seen in the t-th phase… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di. Then, we initialize Θi with Θi−1, and train it using E0:i−1 ∪ Di. The resulting model Θi will be evaluated with a test set containing all classes observed so far. We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches that in each phase it initializes the current model from the prior phase's model and trains on the union of the stored exemplars and the new data, i.e., replaying the exemplars held in its exemplar memory. Training on those stored exemplars corresponds directly to performing incremental learning on the separate storage.),
adjust a number of samples of the first super class and a number of samples of the second super class in the separate storage … (Page 3 Section 3, “an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… M is used to store the exemplars and new coming data as both kinds of data need to be loaded repeatedly during training epochs. In the initial (0-th) phase, data D0, containing the training samples of C0 classes, are used to learn the initial classification model Θ0. In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di… We repeat this training and testing, and report the average accuracy across all phases.”, Page 5 Level-2 Actions, “In the (i−1)-th phase, we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group… Let MA j and MB j denote the memory allocated for the high-entropy and low-entropy groups, respectively, in the j-th phase (j ≤ i): [Equation 2]. Then, we allocate memory evenly to the classes within the group, e.g., if the high-entropy group has 10 classes, each class will have a memory size of 1/10 |M_j|.”, See Figure 1(b) and its Caption,
PNG
media_image1.png
342
471
media_image1.png
Greyscale
”Our proposed method— Reinforced Memory Management (RMM)—is able to learn the optimal and class-specific memory sizes in different incremental phases.”, Figure 2(b) caption, “For the k-th pseudo CIL task, we allocate memory for N times (i.e., in N phases) using the policies pi_n and pi_phi, and compute the cumulative reward R.”
Liu teaches that in each successive phase the Level-2 action re-allocates the memory between the harder and easier groups after the model has been trained in the prior phase, and the algorithm loops back to re-allocate for the next phase.)
and perform incremental learning in a next step on the separate storage ... (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches iterating across all incremental phases, repeating the training on the updated exemplar memory in each successive phase. Re-running the replay training in the following phase corresponds to performing incremental learning in a next step on the separate storage.).
Liu does not teach [a] computer device for fair few-shot shot class-incremental learning (CIL), comprising: memory; and a processor connected to the memory and configured to execute at least one instruction stored in the memory, wherein the processor is configured to construct a separate storage device… when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning
Chowdhury, in the same field of endeavor, teaches [performing an action] when results of the incremental learning satisfy a fairness criterion or when the first accuracy is lower than the second accuracy in the results of the incremental learning (Page 1 Introduction, “To address this problem, we propose a representation learning system– Fairness-aware Incremental Representation Learning (FaIRL). At its core, FaIRL uses an adversarial debiasing setup for removing demographic information by controlling the number of bits (rate-distortion) required to encode the learned representations… We leverage this debiasing setup for incremental learning using an exemplar-based approach, by retaining a small set of representative samples from previous tasks, to prevent forgetting... We propose FaIRL, a representation learning system that learns fair representations, while incrementally learning new tasks, by controlling their rate-distortion function.”, Page 4 Incremental Learning, “To ensure fairness, the system also needs to learn representations that are oblivious to the protected attribute g for both X_new and X_old… This is achieved by minimizing the discriminator loss…”, Page 5 Metrics, “TPR-GAP (De-Arteaga et al. 2019) computes the difference between true positive rates between two protected groups Gap (g, y) = TPR_(g, y) – TPR_(g, y), where g, ¯g are possible values of the protected attribute.”,
Chowdhury teaches an exemplar-based incremental learning system whose objective is fairness across two protected groups, quantified by the true positive rate gap between those groups. The fairness objective of minimizing the discriminator loss corresponds to the fairness criterion used as the criteria to satisfy the fairness of the representations.);
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu's memory allocation method with Chowdhury's fairness objective in order to determine when to adjust the per super class sample counts to reduce the accuracy disparity between the two groups (Page 4 Incremental Learning of Chowdhury).
Liu in view of Chowdhury does not teach [a] computer device for fair few-shot shot class-incremental learning (CIL), comprising: memory; and a processor connected to the memory and configured to execute at least one instruction stored in the memory, wherein the processor is configured to construct a separate storage device… performing incremental learning on the separate storage device
Kumar, in the same field of endeavor teaches [a] computer device for fair few-shot shot class-incremental learning (CIL), comprising: memory; and a processor connected to the memory and configured to execute at least one instruction stored in the memory, wherein the processor is configured to construct a separate storage device… performing incremental learning on the separate storage device (Paragraph 12, “…few-shot incremental learning may be achieved to train a model which accurately classifies for both old and new classes. Catastrophic forgetting of the original classes may be avoided while the addition of new classes is supported. Few-shot incremental learning may be achieved with a smaller training dataset for a new class so that time and resources are saved with respect to annotation/labeling requirements.”, Paragraph 37, “A computer program product … is a term … to describe any set of one, or more, storage media (also called “mediums”) collectively included in a set of one, or more, storage devices that collectively include machine readable code corresponding to instructions and/or data for performing computer operations”, Paragraph 70, “In step 306 of the few-shot incremental deep learning model training process 300 shown in FIG. 3, last hidden state features corresponding to tokens from first and second classes are saved…For example, for a first prototype of a first class, a first set of last hidden state features may be saved and stored in a database… For a second prototype of the second class, another different set, e.g., a fourth set, of last hidden state features may be saved and stored in the database… the database may be within the persistent storage 113 of the client computer”, See Figure 1, Paragraph 40, “PROCESSOR SET 110 includes one, or more, computer processors of any type now known or to be developed in the future... Cache 121 is memory that is located in the processor chip package(s) and is typically used for data or code that should be available for rapid access by the threads or cores running on processor set 110.”
Kumar teaches a few-shot incremental learning that classifies both old and new classes while avoiding catastrophic forgetting using a small new-class dataset, implemented on a computer program product with storage devices. The prototype state features are stored in a database in the computer's storage.)
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury's class-incremental learning with Kumar's few-shot incremental learning implemented on a computer having storage in order to add new classes into a storage device while saving annotation time and resources (Paragraph 12 of Kumar).
Liu in view of Chowdhury and in further view of Kumar does not teach a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy.
Goto, in the same field of endeavor, teaches [identify] a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy (Col. 6 Lines 2-7 of Goto, “For example, if the accuracy of class A is 0.95 while the accuracy of class B is 0.1, the magnitude of augmentation for class B (L.sub.B) may be updated (e.g., L.sub.B=L.sub.B+1) so that the number of training samples of class B, L.sub.BM.sub.B is increased in the next training loop as compared to class A, L.sub.AM.sub.A.”
Goto teaches evaluating class accuracy after training and identifying two classes as a first class having greater accuracy and a second class having less accuracy, expressly comparing the two, e.g., a class A with an accuracy of 0.95 against a class B with an accuracy of 0.1. Goto further teaches increasing the number of training samples of the lower-accuracy second class relative to the higher-accuracy first class in the next training loop.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury in further view of Kumar’s incremental per-group memory allocation with Goto's accuracy difference driven update condition in order to characterize Liu's two exemplar-memory groups by their measured class accuracy (Col. 6 Lines 2-12 of Goto).
Claim 8 recites similar limitations to claims 3 and 4. Therefore, claim 8 is rejected using the same rationale as claims 3 and 4.
Claim 9 recites similar limitations to claim 5. Therefore, claim 9 is rejected using the same rationale as claim 5.
Regarding claim 10,
Liu teaches constructing a separate storage … by using samples of a first super class having first [entropy] and samples of a second super class having second [entropy] lower than the first [entropy] (Page 2 Introduction, “Level-1 function determines how to split memory between the old and new data. Its output action is then inputted into the Level-2 function to determine how to allocate memory for each old class. The overall objective of the function is to maximize the cumulative evaluation accuracy across all incremental phases.”, Page 5 Section 4.1 of Liu, “we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group. Therefore, Level-2 action a[2]i ∈ (0,1) determines how to split memories between harder and easier classes.”
Liu teaches splitting the phase's classes into two groups by training entropy, the higher-entropy difficult classes in one group and the remaining easier classes in the other. These groups are saved into its exemplar-memory partition. The two entropy groups correspond to the two different super classes. The second super class corresponds to Liu's high-entropy difficult group because Liu treats higher entropy as a more difficult class.);
performing incremental learning on the separate storage … (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew. We select Et as representative samples of the data seen in the t-th phase… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di. Then, we initialize Θi with Θi−1, and train it using E0:i−1 ∪ Di. The resulting model Θi will be evaluated with a test set containing all classes observed so far. We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches that in each phase it initializes the current model from the prior phase's model and trains on the union of the stored exemplars and the new data, i.e., replaying the exemplars held in its exemplar memory. Training on those stored exemplars corresponds directly to performing incremental learning on the separate storage.);
adjusting a number of samples of the first super class and a number of samples of the second super class in the separate storage … (Page 3 Section 3, “an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… M is used to store the exemplars and new coming data as both kinds of data need to be loaded repeatedly during training epochs. In the initial (0-th) phase, data D0, containing the training samples of C0 classes, are used to learn the initial classification model Θ0. In the i-th incremental phase, we split M into two dynamic partitions: the exemplar memory Mold and new data memory Mnew… We save E0:i−1 into Mold and free Mnew. Then, we observe new data that contain Ci new classes. We randomly load new data into Mnew until Mnew is full, and all the other new data 3 are discarded. We denote the loaded new data as Di… We repeat this training and testing, and report the average accuracy across all phases.”, Page 5 Level-2 Actions, “In the (i−1)-th phase, we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group… Let MA j and MB j denote the memory allocated for the high-entropy and low-entropy groups, respectively, in the j-th phase (j ≤ i): [Equation 2]. Then, we allocate memory evenly to the classes within the group, e.g., if the high-entropy group has 10 classes, each class will have a memory size of 1/10 |M_j|.”, See Figure 1(b) and its Caption,
PNG
media_image1.png
342
471
media_image1.png
Greyscale
”Our proposed method— Reinforced Memory Management (RMM)—is able to learn the optimal and class-specific memory sizes in different incremental phases.”, Figure 2(b) caption, “For the k-th pseudo CIL task, we allocate memory for N times (i.e., in N phases) using the policies pi_n and pi_phi, and compute the cumulative reward R.”
Liu teaches that in each successive phase the Level-2 action re-allocates the memory between the harder and easier groups after the model has been trained in the prior phase, and the algorithm loops back to re-allocate for the next phase.)
and performing incremental learning in a next step on the separate storage ... (Page 7 Algorithm 1, “for i in 0,...,N do… Observe new data and load Di into Mnew randomly; Initialize Θi with Θi−1 and train it using E0:i−1 ∪Di;”, Page 3 Section 3, “Class-Incremental Learning (CIL) usually assumes (N+1) learning phases: an initial phase and N incremental phases during which the number of classes gradually increases till the maximum… We repeat this training and testing, and report the average accuracy across all phases.”
Liu teaches iterating across all incremental phases, repeating the training on the updated exemplar memory in each successive phase. Re-running the replay training in the following phase corresponds to performing incremental learning in a next step on the separate storage.).
Liu does not teach [a] non-transitory computer-readable recording medium storing at least one program to execute a method for fair few-shot class-incremental learning (CIL) in a computer device, wherein the method for fair few-shot CIL comprises… when results of the incremental learning satisfy a fairness criterion or the first accuracy is lower than the second accuracy in the results of the incremental learning;
Chowdhury, in the same field of endeavor, teaches [performing an action] when results of the incremental learning satisfy a fairness criterion or the first accuracy is lower than the second accuracy in the results of the incremental learning (Page 1 Introduction, “To address this problem, we propose a representation learning system– Fairness-aware Incremental Representation Learning (FaIRL). At its core, FaIRL uses an adversarial debiasing setup for removing demographic information by controlling the number of bits (rate-distortion) required to encode the learned representations… We leverage this debiasing setup for incremental learning using an exemplar-based approach, by retaining a small set of representative samples from previous tasks, to prevent forgetting... We propose FaIRL, a representation learning system that learns fair representations, while incrementally learning new tasks, by controlling their rate-distortion function.”, Page 4 Incremental Learning, “To ensure fairness, the system also needs to learn representations that are oblivious to the protected attribute g for both X_new and X_old… This is achieved by minimizing the discriminator loss…”, Page 5 Metrics, “TPR-GAP (De-Arteaga et al. 2019) computes the difference between true positive rates between two protected groups Gap (g, y) = TPR_(g, y) – TPR_(g, y), where g, ¯g are possible values of the protected attribute.”,
Chowdhury teaches an exemplar-based incremental learning system whose objective is fairness across two protected groups, quantified by the true positive rate gap between those groups. The fairness objective of minimizing the discriminator loss corresponds to the fairness criterion used as the criteria to satisfy the fairness of the representations.);
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu's memory allocation method with Chowdhury's fairness objective in order to determine when to adjust the per super class sample counts to reduce the accuracy disparity between the two groups (Page 4 Incremental Learning of Chowdhury);
Liu in view of Chowdhury does not teach [a] non-transitory computer-readable recording medium storing at least one program to execute a method for fair few-shot class-incremental learning (CIL) in a computer device, wherein the method for fair few-shot CIL comprises… constructing a separate storage device… performing incremental learning on the separate storage device
Kumar, in the same field of endeavor, teaches [a] non-transitory computer-readable recording medium storing at least one program to execute a method for fair few-shot class-incremental learning (CIL) in a computer device, wherein the method for fair few-shot CIL comprises… constructing a separate storage device… performing incremental learning on the separate storage device (Paragraph 12, “…few-shot incremental learning may be achieved to train a model which accurately classifies for both old and new classes. Catastrophic forgetting of the original classes may be avoided while the addition of new classes is supported. Few-shot incremental learning may be achieved with a smaller training dataset for a new class so that time and resources are saved with respect to annotation/labeling requirements.”, Paragraph 37, “A computer program product … is a term … to describe any set of one, or more, storage media (also called “mediums”) collectively included in a set of one, or more, storage devices that collectively include machine readable code corresponding to instructions and/or data for performing computer operations… A computer readable storage medium, as that term is used in the present disclosure, is not to be construed as storage in the form of transitory signals per se”, Paragraph 70, “In step 306 of the few-shot incremental deep learning model training process 300 shown in FIG. 3, last hidden state features corresponding to tokens from first and second classes are saved…For example, for a first prototype of a first class, a first set of last hidden state features may be saved and stored in a database… For a second prototype of the second class, another different set, e.g., a fourth set, of last hidden state features may be saved and stored in the database… the database may be within the persistent storage 113 of the client computer”, See Figure 1.
Kumar teaches a few-shot incremental learning that classifies both old and new classes while avoiding catastrophic forgetting using a small new-class dataset, implemented on a computer program product with storage devices. The prototype state features are stored in a database in the computer's storage.)
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury's class-incremental learning with Kumar's few-shot incremental learning implemented on a computer having storage in order to add new classes into a storage device while saving annotation time and resources (Paragraph 12 of Kumar).
Liu in view of Chowdhury and in further view of Kumar does not teach a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy.
Goto, in the same field of endeavor, teaches [identifying] a first … class having first accuracy and samples of a second … class having second accuracy lower than the first accuracy (Col. 6 Lines 2-7 of Goto, “For example, if the accuracy of class A is 0.95 while the accuracy of class B is 0.1, the magnitude of augmentation for class B (L.sub.B) may be updated (e.g., L.sub.B=L.sub.B+1) so that the number of training samples of class B, L.sub.BM.sub.B is increased in the next training loop as compared to class A, L.sub.AM.sub.A.”
Goto teaches evaluating class accuracy after training and identifying two classes as a first class having greater accuracy and a second class having less accuracy, expressly comparing the two, e.g., a class A with an accuracy of 0.95 against a class B with an accuracy of 0.1. Goto further teaches increasing the number of training samples of the lower-accuracy second class relative to the higher-accuracy first class in the next training loop.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury in further view of Kumar’s incremental per-group memory allocation with Goto's accuracy difference driven update condition in order to characterize Liu's two exemplar-memory groups by their measured class accuracy (Col. 6 Lines 2-12 of Goto).
Claims 2 and 7 are rejected under 35 U.S.C. 103 as being unpatentable over Liu ("RMM: Reinforced Memory Management for Class-Incremental Learning", January 2023), in view of Chowdhury ("Sustaining Fairness via Incremental Learning", 2022), in view of Kumar (US 20240362419 A1), and in further view of Rebuffi ("iCaRL: Incremental Classifier and Representation Learning", 2017).
Regarding claim 2,
Liu teaches storing the samples of the second super class (Page 2 Introduction, “Level-1 function determines how to split memory between the old and new data. Its output action is then inputted into the Level-2 function to determine how to allocate memory for each old class. The overall objective of the function is to maximize the cumulative evaluation accuracy across all incremental phases.”, Page 5 Section 4.1 of Liu, “we split the classes for Di−1 into two groups evenly according to training entropy values: classes with higher values (difficult classes) are in one group and the rest in the other group. Therefore, Level-2 action a[2]i ∈ (0,1) determines how to split memories between harder and easier classes.”
Liu teaches splitting the phase's classes into two groups by training entropy, the higher-entropy difficult classes in one group and the remaining easier classes in the other. These groups are saved into its exemplar-memory partition. The two entropy groups correspond to the two different super classes. The lower-accuracy second super class corresponds to Liu's high-entropy difficult group because Liu treats higher entropy as a more difficult class.)
Liu in view of Chowdhury does not teach the constructing of the separate storage device… the constructing of the separate storage … comprises storing the samples of the second … class by a maximum number that is permitted with respect to the second … class in the separate storage …
Kumar teaches the constructing of the separate storage device (Paragraph 12, “…few-shot incremental learning may be achieved to train a model which accurately classifies for both old and new classes… Few-shot incremental learning may be achieved with a smaller training dataset for a new class so that time and resources are saved with respect to annotation/labeling requirements.”, Paragraph 37, “A computer program product … is a term … to describe any set of one, or more, storage media (also called “mediums”) collectively included in a set of one, or more, storage devices that collectively include machine readable code corresponding to instructions and/or data for performing computer operations…”
Kumar teaches a few-shot incremental learning that classifies both old and new classes while avoiding catastrophic forgetting using a small new-class dataset, implemented on a computer program product with storage devices. The prototype state features are stored in a database in the computer's storage.)
Liu in view of Chowdhury in further view of Kumar does not teach wherein the constructing of the separate storage device comprises storing the samples of the second super class by a maximum number that is permitted with respect to the second super class in the separate storage device.
Rebuffi, in the same field of endeavor, teaches the constructing of the separate storage … comprises storing the samples of the second … class by a maximum number that is permitted with respect to the second … class in the separate storage … (Page 4 Section 2.4, "…when t classes have been observed so far and K is the total number of exemplars that can be stored, iCaRL will use m = K/t exemplars (up to rounding) for each class. By this it is ensured that the available memory budget of K exemplars is always used to full extent, but never exceeded.", "Exemplars p1, . . . , pm are selected and stored iteratively until the target number, m, is met." Rebuffi teaches a bounded exemplar memory in which each class is assigned a target number m of exemplars that represents the largest permitted allocation for that class under the fixed budget K. The construction routine stores exemplars for the class up to that target so that the budget is used to its full extent but never exceeded. Filling the storage for a class up to its permitted target corresponds to storing the samples of the second class by a maximum number that is permitted with respect to the second class in the separate storage.).
Therefore, it would have been obvious before the effective filing date to one of ordinary skill in the art to combine Liu in view of Chowdhury in further view of Kumar's method with Rebuffi's bounded exemplar memory that stores each class up to its permitted target in order to use the fixed memory budget to its full extent and better retain the weaker group (Page 4 Section 2.4 of Rebuffi).
Claim 7 recites similar limitations to claim 2. Therefore, claim 7 is rejected using the same rationale as claim 2.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MAJD MAHER HADDAD whose telephone number is (571)272-2265. The examiner can normally be reached Mon-Friday 8-5 pm.
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, Kamran Afshar, can be reached at (571) 272-7796. 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.
/M.M.H./Examiner, Art Unit 2125
/KAMRAN AFSHAR/Supervisory Patent Examiner, Art Unit 2125
Prosecution Insights