DETAILED ACTION
This Office Action is sent in response to the Applicant’s Communication received on 06/04/2024 for application number 18/733,403. The Office hereby acknowledges receipt of the following and placed of record in file: Specification, Drawings, Abstract, Oath/Declaration, IDS, and Claims.
Claims 1-20 are pending.
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 .
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 3 and 11 recite the limitation "the predetermined number of epochs". There is insufficient antecedent basis for this limitation in the claim.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1, 2, 5-7, 9, 10, 13-15, 17, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Dahiya et al. (Deep Encoders with Auxiliary Parameters for Extreme Classification, published 2023), hereinafter Dahiya, in view of Banerjee et al. (US 20250298839 A1), hereinafter Banerjee, Uy et al. (Deformation-Aware 3D Model Embedding and Retrieval, published 2020), hereinafter Uy, and Xiong et al. (APPROXIMATE NEAREST NEIGHBOR NEGATIVE CONTRASTIVE LEARNING FOR DENSE TEXT RETRIEVAL, published 2020), hereinafter Xiong.
Regarding claim 1, Dahiya teaches,
receiving a plurality of training data-points and a plurality of classifier vectors associated with the training data-points for an extreme classifier model [Sect 3, para 1, The training set is comprised of 𝑁 data points and 𝐿 labels as D… For each data point 𝑖 ∈ [𝑁], its ground truth label vector is y𝑖 ∈ {−1,+1}𝐿, where 𝑦𝑖𝑙 = +1 if label 𝑙 is relevant to the data point 𝑖; Abstract, The paper then proposes a lightweight alternative DEXA that augments encoder training with auxiliary parameters. Incorporating DEXA into existing XC architectures],
each of the classifier vectors mapping a different label of a plurality of labels associated with the extreme classifier model to an embedding space [Sect 3, para 2, We assume that an encoder E𝜃 : X → S𝐷−1 is used to embed data points and labels with 𝜃 being trainable parameters of the encoder. S𝐷−1 denotes the 𝐷-dimensional unit sphere, i.e., the encoder provides 𝐷-dimensional, unit norm embeddings. For simplicity, assume an 1-vs-all-style classifier architecture as W def = {w𝑙}𝑙∈[𝐿] where w𝑙 is the classifier for label 𝑙];
performing a plurality of training epochs [Sect 6, para 4, The Adam optimizer was used to learn the model parameters and its hyperparameters including the number of epochs and learning rate];
sampling a predetermined number of negative labels from a set of negative labels for each of the training data-points [Sect 3, para 4, Instead of training a data point with respect to all its irrelevant labels, most of whom may not even provide useful signals to the model, training is only done with respect to a subset of O (log𝐿) irrelevant labels];
training the encoder and the classifier vectors using the sampled negative labels [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+];
identifying positive labels for each of the training data-points [Sect 4, para 4, To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch]; and
computing a loss based on the sampled negative labels and the identified positive labels for the training data-points; and updating encoder parameters and the classifier vectors based on the computed loss [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+… Note that the above formulation affords (shared) correction terms to all labels. Also, only the hard negative labels and relevant labels of a data point participate in training. To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch].
Dahiya teaches the above limitations of claim 1 including the training data-points (Sect 3, para 1), the training epochs (Sect 6, para 4), and the extreme classifier model (Abstract).
Dahiya does not teach A data processing system comprising: a processor, and a memory storing executable instructions which, when executed by the processor, causes the processor, alone or in combination with other processors, to perform a plurality of functions including: data-points each corresponding to a query, performing training, the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder; wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Banerjee teaches,
A data processing system [Para 0002, system effective to provide query evaluation for image retrieval and conditional image generation] comprising: a processor [Para 0095, The processing element 404 may comprise at least one processor], and a memory storing executable instructions which, when executed by the processor, causes the processor, alone or in combination with other processors, to perform a plurality of functions [Para 0095, The storage element 402 can include one or more different types of memory, data storage, or computer-readable storage media devoted to different purposes within the architecture 400… Different portions of the storage element 402, for example, may be used for program instructions for execution by the processing element 404… the storage element 402 may comprise one or more components of the query understanding and enrichment component 112.] including:
data-points each corresponding to a query [Para 0043, FIG. 1 illustrates an example system 100… in accordance with various aspects of the present disclosure. As shown, a query understanding and enrichment component 112 may receive an input query],
performing training [Para 0045, The discard classifier 114 may be trained to recognize such input queries that have been labeled as “discard” queries in the training data used to train the NLP encoder of query understanding], the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder [Para 0056, As shown in FIG. 1, input queries… may be sent to query encoder; Para 0057, the query embedding (e.g., a vector)].
Banerjee is analogous to the claimed invention as they both relate to encoder-classifier systems that utilize contrastive learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Banerjee and provide using an encoder to generate query embeddings in order to augment classifiers by improving search relevance via semantic understanding.
Dahiya-Banerjee do not teach wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Uy teaches,
wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels [Sect 4, para 3, We use Adam optimizer with a learning rate of 0.001 and train for 350 epochs for all cases; Sect 5, para 3, The Chamfer Distance Triplet (CD-Margin) is trained in the same way with our margin-loss-based approach (Ours-Margin) described in Sec. 3.2; the minibatches are generated with 8 queries and 2 positive and 13 negative random candidates for each of them],
for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels [Sect A.5, For our margin-loss-based approach (Ours-Margin) described in Sec. 3.2, we also tried hard negative mining [51] in the network training. For each query, we generate the set of negative samples Nt with the 8 hardest negatives in Nt (the closest to the query by the learned egocentric distance δ(t;s)) and 5 other randomly selected negatives… The hard negative mining was tested in the fine-tuning, and the network model was first trained in the normal way (with all randomly selected negatives) for 30 epochs].
Uy is analogous to the claimed invention as they both relate to utilizing negative labels. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Uy and provide training with negative labels and hard negative labels in order to [Sect 4, para 4] make training fairly scalable and train on tasks with several millions of labels.
Dahiya-Banerjee-Uy teach the above limitations of claim 1 including the classifier vectors (Dahiya, Sect 3, para 1).
Dahiya-Banerjee-Uy do not teach hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Xiong teaches,
Hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on classifier vectors [Abstract, We then propose Approximate nearest neighbor Negative Contrastive Learning (ANCE), a learning mechanism that selects hard training negatives globally from the entire corpus, using an asynchronously updated ANN index; Sect 2, para 1, Dense Retrieval calculates the retrieval score f() using similarities in a learned embedding space; Sect 4, para 1, we use a simple set up in recent research (Luan et al., 2020) with BERT Siamese/Dual Encoder].
Xiong is analogous to the claimed invention as they both relate to Approximate nearest neighbor methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Xiong and provide Approximate Nearest Neighbor methodologies in order to [Sect 1, para 6] lead to faster learning convergence.
Regarding claim 2, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1 including the ANNS index (Xiong, Abstract).
Uy further teaches,
refreshed once every predetermined number of epochs [Sect A.5, For training efficiency… we cache the latent vectors F(·) and the distance field (PSD) matrices G(·) for all the models in the database and update them every 10 epochs].
Regarding claim 5, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1.
Dahiya further teaches,
wherein the encoder comprises a deep encoder [Title, Deep Encoders with Auxiliary Parameters for Extreme Classification].
Regarding claim 6, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1.
Dahiya further teaches,
wherein per epoch training time is O(log L) where L is a total number of labels used by the extreme classifier model [Sect 3, para 1, Let 𝐿 be the total number of labels in the application; Sect 3, para 4, training is only done with respect to a subset of O (log𝐿)].
Regarding claim 7, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1.
Xiong further teaches,
wherein the loss is binary cross entropy (BCE) loss [Sect 2, para 2, The loss l() can be binary cross entropy (BCE)].
Xiong is analogous to the claimed invention as they both relate to Approximate Nearest Neighbor methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Xiong and provide binary cross entropy loss in order to refine learning models by providing severe penalties for confidently wrong predictions.
Regarding claim 9, Dahiya teaches,
receiving a plurality of training data-points and a plurality of classifier vectors associated with the training data-points for an extreme classifier model [Sect 3, para 1, The training set is comprised of 𝑁 data points and 𝐿 labels as D… For each data point 𝑖 ∈ [𝑁], its ground truth label vector is y𝑖 ∈ {−1,+1}𝐿, where 𝑦𝑖𝑙 = +1 if label 𝑙 is relevant to the data point 𝑖; Abstract, The paper then proposes a lightweight alternative DEXA that augments encoder training with auxiliary parameters. Incorporating DEXA into existing XC architectures],
each of the classifier vectors mapping a different label of a plurality of labels associated with the extreme classifier model to an embedding space [Sect 3, para 2, We assume that an encoder E𝜃 : X → S𝐷−1 is used to embed data points and labels with 𝜃 being trainable parameters of the encoder. S𝐷−1 denotes the 𝐷-dimensional unit sphere, i.e., the encoder provides 𝐷-dimensional, unit norm embeddings. For simplicity, assume an 1-vs-all-style classifier architecture as W def = {w𝑙}𝑙∈[𝐿] where w𝑙 is the classifier for label 𝑙];
performing a plurality of training epochs [Sect 6, para 4, The Adam optimizer was used to learn the model parameters and its hyperparameters including the number of epochs and learning rate];
sampling a predetermined number of negative labels from a set of negative labels for each of the training data-points [Sect 3, para 4, Instead of training a data point with respect to all its irrelevant labels, most of whom may not even provide useful signals to the model, training is only done with respect to a subset of O (log𝐿) irrelevant labels];
training the encoder and the classifier vectors using the sampled negative labels [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+];
identifying positive labels for each of the training data-points [Sect 4, para 4, To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch]; and
computing a loss based on the sampled negative labels and the identified positive labels for the training data-points; and updating encoder parameters and the classifier vectors based on the computed loss [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+… Note that the above formulation affords (shared) correction terms to all labels. Also, only the hard negative labels and relevant labels of a data point participate in training. To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch].
Dahiya teaches the above limitations of claim 9 including the training data-points (Sect 3, para 1), the training epochs (Sect 6, para 4), and the extreme classifier model (Abstract).
Dahiya does not teach A method of training an extreme classifier model, the method comprising: data-points each corresponding to a query, performing training, the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder; wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Banerjee teaches,
A method of training model [Para 0004, FIG. 3 illustrates examples for training a machine learning-based query understanding and enrichment component, in accordance with various aspects of the present disclosure], the method comprising:
data-points each corresponding to a query [Para 0043, FIG. 1 illustrates an example system 100… in accordance with various aspects of the present disclosure. As shown, a query understanding and enrichment component 112 may receive an input query],
performing training [Para 0045, The discard classifier 114 may be trained to recognize such input queries that have been labeled as “discard” queries in the training data used to train the NLP encoder of query understanding], the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder [Para 0056, As shown in FIG. 1, input queries… may be sent to query encoder; Para 0057, the query embedding (e.g., a vector)].
Banerjee is analogous to the claimed invention as they both relate to encoder-classifier systems that utilize contrastive learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Banerjee and provide using an encoder to generate query embeddings in order to augment classifiers by improving search relevance via semantic understanding.
Dahiya-Banerjee do not teach wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Uy teaches,
wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels [Sect 4, para 3, We use Adam optimizer with a learning rate of 0.001 and train for 350 epochs for all cases; Sect 5, para 3, The Chamfer Distance Triplet (CD-Margin) is trained in the same way with our margin-loss-based approach (Ours-Margin) described in Sec. 3.2; the minibatches are generated with 8 queries and 2 positive and 13 negative random candidates for each of them],
for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels [Sect A.5, For our margin-loss-based approach (Ours-Margin) described in Sec. 3.2, we also tried hard negative mining [51] in the network training. For each query, we generate the set of negative samples Nt with the 8 hardest negatives in Nt (the closest to the query by the learned egocentric distance δ(t;s)) and 5 other randomly selected negatives… The hard negative mining was tested in the fine-tuning, and the network model was first trained in the normal way (with all randomly selected negatives) for 30 epochs].
Uy is analogous to the claimed invention as they both relate to utilizing negative labels. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Uy and provide training with negative labels and hard negative labels in order to [Sect 4, para 4] make training fairly scalable and train on tasks with several millions of labels.
Dahiya-Banerjee-Uy teach the above limitations of claim 9 including the classifier vectors (Dahiya, Sect 3, para 1).
Dahiya-Banerjee-Uy do not teach hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Xiong teaches,
Hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on classifier vectors [Abstract, We then propose Approximate nearest neighbor Negative Contrastive Learning (ANCE), a learning mechanism that selects hard training negatives globally from the entire corpus, using an asynchronously updated ANN index; Sect 2, para 1, Dense Retrieval calculates the retrieval score f() using similarities in a learned embedding space; Sect 4, para 1, we use a simple set up in recent research (Luan et al., 2020) with BERT Siamese/Dual Encoder].
Xiong is analogous to the claimed invention as they both relate to Approximate Nearest Neighbor methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Xiong and provide Approximate Nearest Neighbor methodologies in order to [Sect 1, para 6] lead to faster learning convergence.
Method claim 10 recites similar limitations to system claim 2. Therefore, claim 10 is rejected using the same rationale as claim 2.
Method claim 13 recites similar limitations to system claim 5. Therefore, claim 13 is rejected using the same rationale as claim 5.
Method claim 14 recites similar limitations to system claim 6. Therefore, claim 14 is rejected using the same rationale as claim 6.
Method claim 15 recites similar limitations to system claim 7. Therefore, claim 15 is rejected using the same rationale as claim 7.
Regarding claim 17, Dahiya teaches,
receiving a plurality of training data-points and a plurality of classifier vectors associated with the training data-points for an extreme classifier model [Sect 3, para 1, The training set is comprised of 𝑁 data points and 𝐿 labels as D… For each data point 𝑖 ∈ [𝑁], its ground truth label vector is y𝑖 ∈ {−1,+1}𝐿, where 𝑦𝑖𝑙 = +1 if label 𝑙 is relevant to the data point 𝑖; Abstract, The paper then proposes a lightweight alternative DEXA that augments encoder training with auxiliary parameters. Incorporating DEXA into existing XC architectures],
each of the classifier vectors mapping a different label of a plurality of labels associated with the extreme classifier model to an embedding space [Sect 3, para 2, We assume that an encoder E𝜃 : X → S𝐷−1 is used to embed data points and labels with 𝜃 being trainable parameters of the encoder. S𝐷−1 denotes the 𝐷-dimensional unit sphere, i.e., the encoder provides 𝐷-dimensional, unit norm embeddings. For simplicity, assume an 1-vs-all-style classifier architecture as W def = {w𝑙}𝑙∈[𝐿] where w𝑙 is the classifier for label 𝑙];
performing a plurality of training epochs [Sect 6, para 4, The Adam optimizer was used to learn the model parameters and its hyperparameters including the number of epochs and learning rate];
sampling a predetermined number of negative labels from a set of negative labels for each of the training data-points [Sect 3, para 4, Instead of training a data point with respect to all its irrelevant labels, most of whom may not even provide useful signals to the model, training is only done with respect to a subset of O (log𝐿) irrelevant labels];
training the encoder and the classifier vectors using the sampled negative labels [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+];
identifying positive labels for each of the training data-points [Sect 4, para 4, To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch]; and
computing a loss based on the sampled negative labels and the identified positive labels for the training data-points; and updating encoder parameters and the classifier vectors based on the computed loss [Sect 4, para 4, DEXA uses the in-batch-style negative mining strategy proposed in [14] as it had less memory overheads compared to other methods yet offered fast convergence. Encoder training is then done using the following loss function: 𝑁 ∑︁ min {𝜂𝑙 } L({𝜂𝑙}) = 𝑖=1 ∑︁ 𝑙:𝑦𝑖𝑙 =+1 𝑚∈ ˆN𝑖 [𝔑(E𝜃(z𝑚) + a𝐶(𝑘))⊤E𝜃(x𝑖) −𝔑(E𝜃(z𝑙) +a𝐶(𝑙))⊤E𝜃(x𝑖) +𝛾]+… Note that the above formulation affords (shared) correction terms to all labels. Also, only the hard negative labels and relevant labels of a data point participate in training. To further accelerate training, while creating a mini-batch, a single positive label was randomly sampled for each data point in the mini-batch].
Dahiya teaches the above limitations of claim 17 including the training data-points (Sect 3, para 1), the training epochs (Sect 6, para 4), and the extreme classifier model (Abstract).
Dahiya does not teach A non-transitory computer readable medium on which are stored instructions that, when executed, cause a programmable device to perform functions of: data-points each corresponding to a query, performing training, the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder; wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Banerjee teaches,
A non-transitory computer readable medium on which are stored instructions that, when executed, cause a programmable device to perform functions of [The storage element 402 can include one or more different types of memory, data storage, or computer-readable storage media devoted to different purposes within the architecture 400… Different portions of the storage element 402, for example, may be used for program instructions for execution by the processing element 404… the storage element 402 may comprise one or more components of the query understanding and enrichment component 112]:
data-points each corresponding to a query [Para 0043, FIG. 1 illustrates an example system 100… in accordance with various aspects of the present disclosure. As shown, a query understanding and enrichment component 112 may receive an input query],
performing training [Para 0045, The discard classifier 114 may be trained to recognize such input queries that have been labeled as “discard” queries in the training data used to train the NLP encoder of query understanding], the training including: generating query embeddings for each of the data-points that map the data-points to embedding space, the query embeddings being generated using an encoder [Para 0056, As shown in FIG. 1, input queries… may be sent to query encoder; Para 0057, the query embedding (e.g., a vector)].
Banerjee is analogous to the claimed invention as they both relate to encoder-classifier systems that utilize contrastive learning. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Banerjee and provide using an encoder to generate query embeddings in order to augment classifiers by improving search relevance via semantic understanding.
Dahiya-Banerjee do not teach wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels, for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels, and the hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Uy teaches,
wherein: for a first portion of the plurality of training epochs, the sampled negative labels include only uniformly random negative labels [Sect 4, para 3, We use Adam optimizer with a learning rate of 0.001 and train for 350 epochs for all cases; Sect 5, para 3, The Chamfer Distance Triplet (CD-Margin) is trained in the same way with our margin-loss-based approach (Ours-Margin) described in Sec. 3.2; the minibatches are generated with 8 queries and 2 positive and 13 negative random candidates for each of them],
for a second portion of the plurality of training epochs, the sampled negative labels include uniformly random negative labels and hard negative labels [Sect A.5, For our margin-loss-based approach (Ours-Margin) described in Sec. 3.2, we also tried hard negative mining [51] in the network training. For each query, we generate the set of negative samples Nt with the 8 hardest negatives in Nt (the closest to the query by the learned egocentric distance δ(t;s)) and 5 other randomly selected negatives… The hard negative mining was tested in the fine-tuning, and the network model was first trained in the normal way (with all randomly selected negatives) for 30 epochs].
Uy is analogous to the claimed invention as they both relate to utilizing negative labels. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Uy and provide training with negative labels and hard negative labels in order to [Sect 4, para 4] make training fairly scalable and train on tasks with several millions of labels.
Dahiya-Banerjee-Uy teach the above limitations of claim 17 including the classifier vectors (Dahiya, Sect 3, para 1).
Dahiya-Banerjee-Uy do not teach hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on the classifier vectors.
Xiong teaches,
Hard negative labels are identified using an Approximate Nearest Neighbor Search (ANNS) index built on classifier vectors [Abstract, We then propose Approximate nearest neighbor Negative Contrastive Learning (ANCE), a learning mechanism that selects hard training negatives globally from the entire corpus, using an asynchronously updated ANN index; Sect 2, para 1, Dense Retrieval calculates the retrieval score f() using similarities in a learned embedding space; Sect 4, para 1, we use a simple set up in recent research (Luan et al., 2020) with BERT Siamese/Dual Encoder].
Xiong is analogous to the claimed invention as they both relate to Approximate Nearest Neighbor methodologies. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Xiong and provide Approximate Nearest Neighbor methodologies in order to [Sect 1, para 6] lead to faster learning convergence.
Non-transitory computer readable medium claim 18 recites similar limitations to system claim 2. Therefore, claim 18 is rejected using the same rationale as claim 2.
Non-transitory computer readable medium claim 20 recites similar limitations to system claim 6. Therefore, claim 20 is rejected using the same rationale as claim 6.
Claim(s) 3 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over Dahiya in view of Banerjee, Uy, and Xiong, and in further view of Avram et al. (US 20260100267 A1), hereinafter Avram.
Regarding claim 3, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1.
Dahiya-Banerjee-Uy-Xiong do not teach wherein predetermined number of epochs is 5.
Avram teaches,
wherein predetermined number of epochs is 5 [Para 0122, The models are trained with four samples per batch and early stopping is set to five epochs].
Avram is analogous to the claimed invention as they both relate to training deep learning methods. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Avram and provide a predetermined number of epochs in order to manage computational budgets.
Method claim 11 recites similar limitations to system claim 3. Therefore, claim 11 is rejected using the same rationale as claim 3.
Claim(s) 8 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Dahiya in view of Banerjee, Uy, and Xiong, and in further view of Yun et al. (US 20250356123 A1), hereinafter Yun.
Regarding claim 8, Dahiya-Banerjee-Uy-Xiong teach the limitations of claim 1 including the encoder parameters (Dahiya, Sect 4, para 4) and the classifier vectors (Dahiya Sect 3, para 1).
Dahiya-Banerjee-Uy-Xiong do not teach updated using a stochastic gradient descent algorithm.
Yun teaches,
updated using a stochastic gradient descent algorithm [Para 0028, updates the parameters of the pair-comparing model 118 using stochastic gradient descent in combination with back propagation].
Yun is analogous to the claimed invention as they both relate to classifier models. Therefore, it would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified Dahiya’s teachings to incorporate the teachings of Yun and provide stochastic gradient descent update model parameters with high computational efficiency.
Method claim 16 recites similar limitations to system claim 8. Therefore, claim 16 is rejected using the same rationale as claim 8.
ALLOWABLE SUBJECT MATTER
Claims 4, 12, and 19 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SYED RAYHAN AHMED whose telephone number is (571)270-0286. The examiner can normally be reached Mon-Fri ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, David Yi can be reached at (571) 270-7519. 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.
/SYED RAYHAN AHMED/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights