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 .
DETAILED ACTION
The instant application having Application No. 18/896,700 is presented for examination by the examiner. Claims 1, 3, 4, 10, 11, 14, 15, and 19 are amended, claims 1-20 have been examined.
Response to Arguments
Applicant' s arguments filed 06/24/2026, with respect to the rejection(s) of claim(s) 1, 10 and 19 under 35 U.S.C. 103 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
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 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 of this title, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 4, 8, 10, 16-17 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in further view of Zhu (US 20220230625 A1).
Regarding Claim 1
Martin discloses:
A method comprising: receiving a request for an unlearning of a content from a machine learning (ML) model trained using training data including the content, wherein the unlearning reconfigures the ML model to be trained using the training data independent of the content (Martin ¶¶0013, 0017, 0069: teaches receiving a request for unlearning content from a machine learning model trained using training data including the content, where the service receives a description of a concept to be unlearned from a pre-trained ML model and generate an adapter that modifies model parameters to cause the model to forget the concept.);
performing a content detection check of the ML model for the content based on at least one of the training data, an output by the ML model, or a source code file for the ML model (Martin ¶¶0060, 0064: teaches performing a content detection check of the ML model for the content based on training data and model outputs, where the model unlearning training service processes negative and positive prompts derived from training data to identify activation regions most sensitive to the concept to be forgotten.);
identifying, from a graph representation of at least a portion of the ML model, one or more nodes of the ML model associated with the content based on the corresponding concepts mapped to the content (Martin ¶¶0060–0062 and 0064 teaches identifying nodes of the ML model associated with the content. A neural network is a computational graph in which neurons correspond to nodes and weighted connections correspond to edges. The model unlearning training service processes prompts through the pre-trained model and captures activation volume maps representing neuron activations across layers of the neural network. A statistical profiling engine analyzes the activation distributions to determine which neurons contain information indicative of the concept to be forgotten and gradient analysis identifies activation regions most sensitive to the concept, thereby identifying neurons of the model associated with the concept.); and
performing a selective parameter dampening of the one or more nodes (Martin ¶¶0048 and 0065–0067: teaches performing a selective parameter dampening of nodes of the ML model, where when a detector identifies activations associated with a concept to be forgotten, parameter updates are selectively applied to weights of neurons receiving those activations, including setting weights to zero, adding noise to fuzz activations, or scaling the weights to steer the model away from the concept.).
Martin teaches performing a content detection check of the ML model for the content based on training data and model outputs, where prompts associated with the concept to be forgotten are processed by the model and activation volume maps are captured to determine neuron activations associated with the concept. Martin does no disclose accessing a knowledge graph from the content, wherein each node corresponds to one of the corresponding concepts from the content and each edge represents an association between two or more of the corresponding concepts; converting the knowledge graph to graph embedding in a vector space; projecting outputs of the ML model into the vector space; and performing an overlap detection between the graph embeddings and the projected outputs to identify the corresponding concepts. However, Jiang teaches deriving a knowledge graph from content by extracting entities and corresponding relationships directly from electronic documents, wherein nodes of the knowledge graph correspond to the extracted entities and edge correspond to the extracted relationships (Jiang ¶¶0003, 0058, 0060-0061). Jiang further teaches performing knowledge graph embedding to map the entities and relationships of the knowledge graph into vectors within a lower dimensional vector space (Jiang ¶¶0058, 0062).
It would have been obvious to a POSTIA before the effective filling date of the claimed invention to modify Martin’s concept localization process according to Jiang. Marting identifies activations associated with content to be unlearned but does not provide a content derived semantic representation for determining the concepts represented by those activations. Jiang teaches deriving such a representation directly from content by extracting entities and relationships, representing them as a knowledge graph, and embedding the graph entities and relationships in a vector space. A POSTIA therefore would have applied Jiang’s technique to Martin’s content to obtain a structured, vectorized representation of the concepts associated with the content to be unlearned.
Martin/Jiang does not teach projecting outputs projecting outputs of the ML model into the vector space containing the knowledge graph embedding and performing an overlap detection between the graph embeddings and the projected outputs to identify the concept in the ML model that is mapped to the content. However, Zhu teaches generating structure aware entity embeddings from a knowledge graph using a graph convolution network and projecting machine learning model output embeddings and knowledge-based relation embeddings into a shared semantic space (Zhu ¶¶0069-0070, 0077-0078, 0126-0129). Zhu teaches the integration of the language module and knowledge module comprises “projecting entity and relations output embeddings and language text embeddings into a shared semantic space” (Zhu ¶¶0128). Zhu further teaches identifying a corresponding entity based on proximity in the embedding space, wherein the model predicts an entity whose embedding is closet to the embedding associated with an entity mention (Zhu ¶¶0082-0083).
It would have been obvious for a POSITA before the effective filing date of the claimed invention to further modify the Martin/Jiang combination according to Zhu. Zhu teaches projecting knowledge-based relation embeddings and Machine Learning model output embedding into a shared semantic space and identifying corresponding semantic entities based on proximity in that space. A POSITA would therefore have applied Zhu’s shared space alignment technique to Jiang’s content derived knowledge graph embeddings and outputs/activations representations associated with Martin’s ML model so that concepts represented by Martin’s detected activations could be identified based on proximity to the content derived graph embeddings. The combination would predictable provide a semantically grounded identification of model representations corresponding to the content before Martin selectively dampens the associated model parameters.
Regarding Claim 4
Martin discloses:
The method of claim 1, wherein the content comprises one of a copyrighted content or a privacy protected content, and wherein the ML model comprises one of a neural network (NN) having one or more neurons corresponding to the one or more nodes that activate based on the relevant concepts learned from the content or a large language model (LLM) that provides responses based on a knowledge base including the content (Martin ¶¶0016, ¶¶0018, ¶0047, ¶0054, ¶0062-0063: teaches that the content comprises trademark (copyrighted) content learned from training data, and that the ML model comprises a neural network having neurons that activate based on relevant concepts learned from that content and a LLM that models semantic relationships based on a knowledge base derived from training data.).
Regarding Claim 8
Martin discloses:
The method of claim 1, further comprising: evaluating a performance of the ML model after the performing the selective parameter dampening to a base performance of the ML model prior to the performing the selective parameter dampening, wherein the evaluating includes testing at least one benchmark ML task performed by the ML model (Martin ¶¶0061–0063 and 0066–0067: teaches evaluating a performance of the ML model after performing selective parameter dampening, where model parameters are selectively modified (e.g., zeroing weights, adding noise, or scaling weights) to unlearn a concept, and the modified model is evaluated by processing structured prompt datasets and capturing activation outputs to compare behavior between positive and negative samples using statistical metrics including false positive and false negative rates).
Regarding Claim 10
Claim 10 is directed to a system corresponding to the method of claim 1. Claim 10 recites substantially the same limitations as claim 1 in system form and is therefore rejected under the same rationale set forth above for claim 1.
Regarding Claim 16
Martin discloses:
The system of claim 10, wherein performing the selective parameter dampening comprises: representing the ML model as a graph having a plurality of nodes connected by a plurality of edges, wherein the plurality of nodes correspond to neurons of the ML model and the plurality of edges correspond to synapses of the ML model (Martin ¶¶0065–0066 teaches a neural network having neurons connected by weighted inputs and outputs across layers, and a neural network is a computational graph in which neurons correspond to nodes and weighted connections correspond to edges.); and analyzing one or more activation patterns and one or more weight distributions of the ML model using the graph during the execution of the ML model (Martin ¶¶0061–0064 teaches analyzing activation patterns by capturing activation volume maps and comparing neuron activation distributions across samples, and ¶¶0065–0066 teaches analyzing weight distributions by adjusting weights associated with activations during model execution.).
Regarding Claim 17
The system of claim 16, wherein performing the selective parameter dampening further comprises: creating a negation vector based on the analyzing and the graph for the selective parameter dampening (Martin ¶¶0017 and 0066: teaches creating a representation based on analyzed activation patterns (“fingerprint”) of a concept and applying parameter updates that steer the model away from the concept, and modifying weights in a neural network constitutes applying a vector in parameter space, thereby corresponding to creating a negation vector for selective parameter dampening.).
Regarding Claim 19
Claim 19 is directed to a non-transitory machine-readable medium corresponding to the method of claim 1. Claim 19 recites substantially the same limitations as claim 1 in a computer-readable medium form and is therefore rejected under the same rationale set forth above for claim 1.
Regarding Claim 20
Claim 20 is directed to a non-transitory machine-readable medium corresponding to the system of claim 17. Claim 20 recites substantially the same limitations as claim 17 in a computer-readable medium form and is therefore rejected under the same rationale set forth above for claim 17.
Claims 2-3 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 1 and 10 above, and in further view of Hakim (US 20230421562 A1).
Regarding Claim 2
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach verifying a requestor of the request based on a requestor identifier and a verification token received with the request prior to performing the content detection check, including checking user behavior associated with the requestor identifier against a user record and authorizing the verification token. Hakim teaches verifying a requestor of a request based on a requestor identifier and a verification token received with the request. Specifically, Hakim discloses that a client authenticates through an identity provider and receives an access token that is included with subsequent API requests, and that the gateway device validates the access token to determine whether the client is authorized to access the requested resource (Hakim ¶¶0007, 0058–0061). Hakim further teaches checking behavior and other information associated with the requestor identifier against stored records, including comparing client account identifiers and IP addresses indicating accounts or devices associated with harmful activity (Hakim ¶¶0063, 0071). Hakim additionally teaches authorizing the token before permitting the request to proceed to the requested resource and otherwise blocking or quarantining the request if the authorization or risk determination fails (Hakim ¶¶0061–0063, 0067–0070).
Hakim performs authentication checks using requestor identifiers, access tokens, and stored behavioral information before allowing a request to access system resources. Martin, Jiang and Zhu already process requests that trigger machine unlearning operations, and incorporating requestor verification as taught by Hakim would ensure that only authenticated and authorized requestors are permitted to initiate those operations. The claim is obvious because one of ordinary skill in the art would recognize that applying known authentication and authorization techniques to a system performing machine unlearning operations would improve system security and reliability by preventing unauthorized requests from triggering model processing.
Regarding Claim 3
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach performing contextual verification of the requestor based on external verification sources and requester privileges when verifying the requestor of the request. Hakim teaches performing contextual verification based on external verification sources and requester privileges when verifying a requestor. Specifically, Hakim discloses that a client authenticates through an external identity provider and an authentication platform which verifies the client’s identity and issues authentication tokens that are used to validate the client in subsequent requests (Hakim ¶¶0007, 0058–0061). Hakim further teaches determining a scope or level of authorization associated with the client based on the authentication process and access token, where the token indicates which resources the client is permitted to access and what actions the client is authorized to perform (Hakim ¶¶0060–0061). These authentication services and authorization scopes constitute contextual verification using external verification sources and requester privileges.
Hakim performs contextual verification using external identity providers and authorization scopes associated with authentication tokens when determining whether a request should be permitted. Martin, Jiang and Zhu already process requests that initiate machine unlearning operations, and incorporating contextual verification as taught by Hakim would ensure that the system evaluates externally verified identity information and requester privileges before performing those operations. The claim is obvious because one of ordinary skill in the art would recognize that applying known contextual authentication and authorization techniques to systems performing machine unlearning model operations would improve system security, producing the predictable result of restricting such operations to properly verified and authorized requestors.
Regarding Claim 13
Claim 13 is directed to a system corresponding to the method of claim 2 and recites substantially the same limitations in system form. Accordingly, claim 13 is rejected under the same rationale set forth above for claim 2, for the same reasons discussed above.
Claims 5 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 1 above, and in further view of Padgett (US 20240160902 A1).
Regarding Claim 5
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach performing the content detection check by comparing output of the ML model to a database of copyrighted content using a content similarity detection operation and flagging matches of the output to the copyrighted content. Padgett teaches performing a content detection check by comparing output generated by a generative AI model to a repository of pre-existing content using a similarity assessment layer that determines a similarity measure for the output content with respect to the repository and compares that similarity measure to a threshold value (Padgett ¶¶0016, 0018, 0024, 0064, 0077–0083). Padgett further teaches that the pre-existing content can include content subject to copyright protection and that the similarity assessment layer can determine when the output is too similar to such pre-existing content by finding a match within a threshold value (Padgett ¶¶0021, 0034, 0037–0038, 0100–0102).
Padgett performs a content comparison analysis using similarity detection against repositories of pre-existing protected content. Martin, Jiang and Zhu already process model content and identify concepts associated with output representations, and incorporating Padgett’s similarity assessment comparison would improve detection of outputs that contain copyrighted material. The claim is obvious because one of ordinary skill in the art would recognize that combining known similarity content comparison techniques with systems performing machine learning model unlearning would improve identification of copyrighted, producing the predictable result of flagging outputs that should be subject to further review or corrective action.
Claims 6 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 1 above, and in further view of Shama (US 20220300640 A1).
Regarding Claim 6
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach performing the content detection check by identifying privacy protected data in the training data, determining whether the privacy protected data was masked during training of the ML model, and determining, when the privacy protected data was unmasked during training, a contribution of the privacy protected data to the training of the ML model. Shama teaches identifying sensitive (privacy protected) data fields within a dataset used to train a model, including annotating and identifying private information in the training data (¶¶0011, 0015, 0019, 0028). Shama further teaches applying and evaluating privacy preserving techniques, including masking, thereby determining whether such data was masked during training (¶¶0013, 0016, 0030). Shama additionally teaches determining feature weight using techniques such as regression coefficients and evaluating the impact of such features on model accuracy, which corresponds to determining a contribution of the privacy protected data to model training (¶¶0020, 0029, 0034).
Shama performs a privacy focused analysis on training data by identifying sensitive data, evaluating whether masking is applied, and determining the importance of such data to model training. Martin, Jiang and Zhu already analyze and modify model behavior based on identified content, and incorporating Shama’s techniques would improve identification and evaluation of privacy training data. The claim is obvious because one of ordinary skill in the art would recognize that applying known privacy analysis and feature importance techniques to a system performing machine learning model unlearning would improve assessment of how privacy protected data contributes to the model, producing the predictable result of more accurately determining whether such data should be addressed.
Claims 7 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 1 above, and in further view of Kim (US 11423142 B2) and Carlini1 N.P.L. (The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks).
Regarding Claim 7
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach generating, from the source code file, a credential dependency graph of credentials learned during a training of the ML model; and determining a score representing whether one of the credentials corresponding to the content is capable of being leaked by the ML model. Kim teaches generating a dependency graph from source code by constructing control and data dependency graphs in which nodes represent program statements or data objects and edges represent dependencies between them (Column 4, Lines 39-64). Kim further teaches identifying confidential data and model parameters determined during training and analyzing their dependencies within the graph to determine how sensitive information propagates through the program (Column 4, Line 39 – Column 5, Line 20 and Column 7, Lines 4-28).
Kim teaches a structural dependency analysis over sensitive data and model related information derived from machine learning systems. Martin, Jiang and Zhu already identify content and corresponding model components, and incorporating Kim’s dependency graph generation would improve those systems by enabling structured identification of relationships between sensitive credential data and model components. The claim is obvious because one of ordinary skill in the art would recognize that applying known dependency graph analysis techniques to machine learning systems would improve tracking and isolation of sensitive data learned during training, producing the predictable result of more effectively identifying and managing credential-related information within the model.
Martin, Jiang, Zhu and Kim collectively teach identifying and removing content from a machine learning model by detecting concept associated activations and generating a dependency graph from source code in embedding space for unlearning. However Martin, Jiang, Zhu and Kim do not explicitly teach determining a score representing whether one of the credentials corresponding to the content is capable of being leaked by the ML model. Carlini teaches determining such a score, where Carlini discloses a quantitative metric of exposure” that measures a model’s potential for unintended memorization of unique or rare sequences in the training data including sensitive information such as credit card numbers and other secret sequences (Page 1). Carlini further teaches that this exposure metric characterizes the likelihood of training-data disclosure and can be used to assess whether such secrets can be extracted from the model, thereby representing whether the credentials are capable of being leaked by the ML model (page 3).
Carlini teaches a leakage risk evaluation for sensitive content represented in a machine learning model. Martin, Jiang, Zhu and Kim already identify content and corresponding model components for unlearning, and incorporating Carlini’s exposure metric would allow those systems to quantify whether sensitive credentials associated with the identified content remain vulnerable to disclosure. The claim is obvious because one of ordinary skill in the art would recognize that applying known quantitative leakage-risk scoring techniques to a system performing machine learning model unlearning would improve evaluation of whether sensitive credential information is still memorized and extractable, producing the predictable result of more reliably assessing and mitigating unintended data leakage from the model.
Claims 9 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 1 above as applied to claim 1 and 10 above, and in further view of Dialani (US 20250190784 A1).
Regarding Claim 9
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach generating an unlearning proof after performing the selective parameter dampening based on evaluating model performance and responding to a request with that proof. Dialani teaches generating proof artifacts in connection with a machine unlearning process. Dialani teaches that upon receiving a deletion request for data to be unlearned, a proof subsystem generate a hash value corresponding to intermediate and final model instances resulting from stages of the unlearning process and stores those values for later verification (¶¶0018-0019, 0031-0033, 0090-0091). Dialani also teaches evaluating whether unlearning was successfully completed through direct testing and through statistical analysis showing a reduction in the occurrence of the target information (¶¶0021). Dialani also teaches responding to a verification request by sending a stored proof artifact corresponding to the resulting unlearned model (¶¶0106).
Dialani teaches a post unlearning verification which generates proof tied to the resulting model after the unlearning operation and providing that proof in a response request. Martin, Jiang and Zhu already teach an unlearning operation, and incorporating the teaching of Dialani proof generation and verification technique would have enabled confirmation that the targeted concept had in face been removed from the model. The claim would have been obvious because one of ordinary skill in the art would recognized that applying known proof generation and verification techniques to a machine unlearning operation would improve reliability and compliance by providing verifiable evidence that the unlearning operation was successful.
Regarding Claim 12
Claim 12 is directed to a system corresponding to the method of claim 9 and recites substantially the same limitations in system form. Accordingly, claim 12 is rejected under the same rationale set forth above for claim 9, for the same reasons discussed above.
Claims 11 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 10 above, and in further view of Headley (US 20250184360 A1).
Regarding Claim 11
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach that the content comprises privacy protected data for a user and that such privacy protected data is identified using a sensitive data detection component and a database storing flagged instances of sensitive user data. Headley teaches identifying sensitive data using DLP functionality based on fingerprints of sensitive files and maintaining those fingerprints in a fingerprint database. Specifically, Headley discloses that cloud services may store files containing sensitive data, thereby making them useful for generation of DLP fingerprints (¶¶0032), that a CASB platform may monitor cloud storage and maintain a set of DLP fingerprints in a fingerprint database (¶¶0033), and that a security enforcement point may perform DLP scanning using the stored DLP fingerprints to identify sensitive content (¶¶0034, 0038–0039, 0051–0053).
It would have been obvious to a person of ordinary skill in the art at the time of the invention to use Headley’s sensitive data detection and fingerprint database techniques in the unlearning framework of Martin, Jiang and Zhu in order to identify privacy protected user data that should be targeted for removal, thereby improving the system’s ability to detect protected user content using stored flagged instances of sensitive data before performing unlearning operations.
Regarding Claim 18
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However Martin, Jiang and Zhu do not explicitly teach that the content comprises protected content identified using a retrieval engine and a repository of protected contents utilized as a benchmark. On the other hand, Headley teaches identifying protected (sensitive) content using a repository of stored content representations. Specifically, Headley discloses maintaining a fingerprint database of sensitive data (DLP fingerprints) generated from files stored in cloud services and using that database to identify protected content by performing fingerprint matching against files (¶¶0033–0034, 0038–0039, 0049–0053). Headley further teaches obtaining and accessing these fingerprints from storage systems and using them during scanning operations, which corresponds to retrieving stored representations of protected content for comparison, i.e., a retrieval engine accessing a repository of protected content for identification purposes.
Headley teaches a comparison identification of protected content using stored reference data. Martin, Jiang and Zhu already identifies content within a model for further unlearning operation, and incorporating the repository retrieval and matching techniques of Headley would allow the system to more accurately and reliably identify protected content using known reference data. The claim is obvious because one of ordinary skill in the art would recognize that combining model content identification with database fingerprint matching would improve detection accuracy and security by ensuring that protected content is identified based on known benchmarks, producing the predictable result of more reliable identification of protected content.
Claims 14 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1) as applied to claim 10 above, and in further view of TOMBERG (US 20200311115 A1).
Regarding Claim 14
Martin, Jiang and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang and Zhu do not explicitly teach generating a mapping of a plurality of concepts where nodes correspond to concepts and edges correspond to associations between the concepts. On the other hand, Tomberg teaches representing concepts in a structured graph form, where concepts are organized as nodes in a taxonomy and relationships between concepts are represented as edges (¶¶0060). Tomberg further teaches mapping text phrases to concepts within this taxonomy using vector representations and embedding mapping functions, thereby correlating input content with concepts in the structured concept graph (¶¶0053, 0062). Additionally, Tomberg teaches that relationships between concepts capture semantic and contextual associations, including hierarchical and attribute relationships within the taxonomy (Tomberg ¶¶0048, 0060).
Tomberg performs a concept organization and relationship modeling between content and learned concepts using a graph representation. Martin, Jiang and Zhu already teach identity concepts within the model via activation analysis, and organizing those concepts into a structured mapping of nodes and edges as taught by Tomberg would improve identification and correlation of concepts associated with the content. The claim is obvious because one of ordinary skill in the art could combine known techniques for activation concept identification with graph representations of concept relationships to better structure and correlate learned concepts, producing the predictable result of more effectively identifying and locating model components associated with the content.
Claims 15 are rejected under 35 U.S.C. 103 as being unpatentable over Martin (US 20250103877 A1), in view of Jiang (US 20200125648 A1), in view of Zhu (US 20220230625 A1), in view of TOMBERG (US 20200311115 A1) as applied to claim 14 above, and in further view of MINERVINI (US 20180144252 A1).
Regarding Claim 15
Martin, Jiang, and Zhu collectively teach identifying and removing concepts from a machine learning model by analyzing model activations associated with the concept and mapping related content in a vector embedding space to locate semantically related representations for machine unlearning. However, Martin, Jiang, and Zhu are silent in explicitly teaching generating a knowledge graph of concepts in a vector space, projecting outputs of the ML model into that vector space, and determining overlaps between the knowledge graph and the outputs to identify nodes. On the other hand, Tomberg teaches generating a structured knowledge graph of concepts, where concepts are represented as nodes and relationships between concepts are represented as edges within a taxonomy (¶¶0060). Tomberg further teaches representing these concepts using node embeddings in a vector space and mapping input content into that vector space to identify corresponding concepts based on proximity and similarity (¶¶0049–¶0050, ¶0062). Martin, Jiang, Zhu and Tomberg can be combined to teach generating a knowledge graph of the plurality of concepts in the vector space but combined they do not teach determining one or more overlaps between the knowledge graph and the outputs, wherein the node is identified based at least on the one or more overlaps. Minervini teaches embedding entities and relationships of a knowledge graph into a continuous vector space and computing similarity and scoring functions between embedded representations to determine relationships and likelihoods between elements, including comparing vector representations using distance metrics and similarity measures (¶¶0045–¶0047, 0060–0063). Minervini further teaches that such vector comparisons enable determining correspondences between inputs and elements of the knowledge graph based on similarity in the embedding space. Tomberg and Minervini perform the same type of mapping and comparison between content representations and graph concept representations using vector embeddings and similarity computations.
Martin and Jiang already analyzes model outputs associated with the content, and projecting those outputs into a vector space and comparing them with a knowledge graph as taught by Tomberg and Minervini would improve identification of relevant concepts associated with the content. The claim is obvious because one of ordinary skill in the art could combine known techniques for activation output analysis with graph concept representations and vector similarity comparisons to identify corresponding nodes, producing the predictable result of more accurately locating concepts in the model that correspond to the content.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any extension fee pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SAAD A ABDULLAH whose telephone number is (571) 272-1531. The examiner can normally be reached on Monday - Friday, 8:30am - 5:00pm, EST. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Lynn Feild can be reached on (571) 272-2092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/SAAD AHMAD ABDULLAH/Examiner, Art Unit 2431
/SHIN-HON (ERIC) CHEN/Primary Examiner, Art Unit 2431
