DETAILED ACTION
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 .
This action is responsive to the filing on 05/19/2025. Claims 1, 3, 5, 6-9, 15-16 and 19-23 have been amended. Claims 1, 3, 5-9, 11, 15-17, 19-26 are pending in this case. Claims 25-26 are newly added. Claims 2, 4, 10, 12-14 and 18 are cancelled.
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 05/19/2025 has been entered.
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.
Claims 1, 3, 5, 7-9, 11, 16-17 and 19-26 are rejected under 35 U.S.C 103 as being unpatentable over “Learning to Generalize: Meta-Learning for Domain Generalization”, https://arxiv.org/pdf/1710.03463, Li et al., 2017, hereinafter referred to as Li in view of “Learning without Forgetting”, https://arxiv.org/pdf/1606.09282, Hoiem et al, 2017, hereinafter referred to as Hoiem and further in view of Albright et al. (US Pub No.: 20190130218 A1), hereinafter referred to as Albright.
With respect to claim 1, Li disclose:
A computer-implemented method for training a neural network model for sequentially learning a plurality of domains associated with a task, the plurality of domains comprising a < multiple source domains >, the computer-implemented method comprising: (On Pages 2–3 (Meta-Learning Domain Generalization), Li teaches training a machine-learning model using a plurality of domains associated with a classification task. In particular, Li discloses a domain-generalization framework in which training data are obtained from multiple source domains and the sources are divided into meta-train and meta-test domains to simulate domain shifts during training, thereby training the model to generalize across different domains.)
Determining at least one set of auxiliary model parameters for the neural network model by simulating, using a processor, at least one first optimization step based on a set of current model parameters for the neural network model and at least one auxiliary domain, wherein the at least one auxiliary domain is associated with a primary domain comprising data points for training the neural network model (On page 3, Li teaches determining an updated set of model parameters by simulating an optimization step based on the current model parameters and training data associated with a domain. In particular, Li's meta-learning procedure performs a simulated optimization of the current model parameters using a meta-training domain to obtain updated model parameters, which are subsequently utilized in the meta-optimization procedure.)
Wherein the at least one set of auxiliary model parameters minimizes a loss associated with a respective auxiliary domain of the at least one auxiliary domain with respect to the current model parameters (On page 3 (Algorithm 1 & 2), Li teaches determining updated model parameters from current model parameters through a gradient-based optimization step and evaluating a loss G on a respective meta-test domain using the updated parameters. )
Determining a set of primary model parameters for the neural network model by performing, using the processor, a second optimization for the neural network model comprising minimizing a combined loss function, wherein the combined loss function combines: (On page 3, Li teaches a two-level optimization. First, it performs an inner/meta-train optimization from the current parameter, producing an intermediate parameter. Then Li constructs the overall meta-objective, where F represents the meta-training loss and G represents the meta-test loss evaluated after the simulated update.)
A first loss function associated with the set of current model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a first loss F(Θ) representing the meta-training loss using current parameters.)
A second loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a second loss G(Θ), using the updated/intermediate parameter and combining.)
Updating, using the processor, the neural network model with the set of primary model parameters (On page 3, Li teaches updating the model. The model is parameterized by Θ, so the gradient of Θ calculated with respect to this loss function is ∇Θ, and optimization will update the model as Θ=Θ−α∇Θ.)
With respect to claim 1, Li does not explicitly disclose:
<A first primary domain and a second primary domain>
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
Each of a first stage of training and a second stage of training
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data
However, it is known by Hoiem to disclose:
<A first primary domain and a second primary domain> known as multiple source domains (Fig.3 & Page 5 (Relationship to joint training), Hoiem discloses original/old-task training dataset new task dataset.)
Each of a first stage of training and a second stage of training (Fig.3 & Page 5 (Relationship to joint training & efficiency comparison). Hoiem teaches by gradually adding new capabilities to an existing CNN when the training data for the existing capabilities are unavailable. Prior/original training that established CNN's existing capabilities.)
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain (On page 5, Hoiem teaches a convolutional neural network previously trained using original-task training data, thereby teaching a first stage of training in which the neural network is trained on data points of an original/first domain. Hoiem subsequently trains the network using only new-task data while the original-task training data are unavailable. )
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain (On page 4 (3 Learning without forgetting), Hoiem subsequently teaches subsequently training an already-trained convolutional neural network using data associated with a new task (i.e., a secondary primary domain), wherein the training data associated with the network's previously learned capabilities (i.e., data points of a first primary domain) are unavailable. Hoiem further teaches learning without forgetting and uses only the new-task data to train the network while preserving the previously learned capabilities, thereby teaching subsequent training in the second primary domain by accessing data points of the first primary domain.)
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task (The examiner selects: classification task: On page 2 (1–2 paragraphs), Hoiem teaches updating an existing convolutional neural network using new-task training data while preserving previously learned capabilities, wherein the resulting updated convolutional neural network performs visual classification tasks associated with the new and previously learned capabilities, thereby teaching an updates neural network model used to perform as least a classification task.)
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data (The examiner selects: image. On page 3 and Fig. 2, Hoiem teaches that data points of the training domains comprise image data, as the disclosed convolutional neural network is trained and evaluated using image datasets for computer-vision/classification tasks.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to train multiple source domains as taught by Li, while using new-task training data while attempting to preserve the network's existing capabilities as taught by Hoiem. The motivation for doing so would have been to improve training domain performance should also improve testing domain performance (See Page 1 of the Abstract of Li.)
With respect to claim 1, Li in view of Hoiem does not explicitly disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
However, it is known by Albright to disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation (In paragraphs [0033-0034], Albright teaches generating augmented training data by modifying images of an initial training dataset, wherein one or more transformations are performed on the initial training images to generate modified training images. The disclosed transformations include scaling, rotation, translation, brightness modification, lighting modification, and perspective modification. Thus, the Albright reference teaches or at least suggests selecting from a plurality of available transformations and manipulating data points of an original training domain using the selected transformation to generate modified training data.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, with one or more transformations on the portion of the image as taught by Albright. The motivation for doing so would have been to replace fine-tuning with similar old and new task datasets for improved new task performance (See (Abstract) of Hoiem.)
Regarding claim 3, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim [[2]] 1, wherein the data manipulation is performed automatically using the processor (On page 3, Li teaches performing the disclosed data manipulation as part of an automated neural-network training algorithm executed by a computing system. Accordingly, a person of ordinary skill in the art would have understood that the algorithmic manipulation of the training data is performed automatically using a processor.)
Regarding claim 5, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Albright disclose:
The computer-implemented method of claim [[2]]1, wherein the modifying the one or more data points of the primary domain via data manipulation comprises randomly selecting, using the processor, one or more transformations from [[a]]the set of transformations (In paragraph [0037], Albright teaches parameter determination module 275 determines one or more transformation parameters based on several factors. Each transformation parameter determines the amount of the transformation performed by the transformation module.)
Regarding claim 7, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim 1, wherein the second loss function has an objective of avoiding catastrophic forgetting and wherein the combined loss function further combines a third loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the at least one auxiliary domain, wherein the third loss function has an objective of encouraging domain adaptation (On page 3, Li teaches minimizing a combined optimization objective comprising a first loss F(Θ) associated with meta-training domains and a second loss G(Θ) associated with virtual meta-test domains, wherein the losses are combined as F(Θ)+βG(Θ−αF(Θ)) to determine/update the model parameters.)
Regarding claim 8, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim [[1]]7, wherein the first loss function has an objective associated with task learning (On page 3, Li teaches a first loss F(Θ) representing the meta-training loss using current parameters.)
Regarding claim 9, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim 8, wherein the second optimization uses via gradient descent (On page 3, Li teaches to perform optimization of the neural network model via gradient descent, wherein model parameters are updated based on the gradient of a loss, thereby teaching wherein the optimization uses gradient descent.)
Regarding claim 11, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Hoiem disclose:
The computer-implemented method of claim 1, further comprising: initializing, using the processor, the neural network model, wherein initializing the neural network model comprises setting model parameters of a pre-trained neural network model as initial model parameters for the neural network model to fine-tune the pre-trained neural network model; wherein said fine-tuning the pre-trained neural network model comprises performing the second optimization (On page 2, Hoiem teaches initializing a neural network using an existing previously trained convolutional neural network having previously learned model parameters and subsequently training/adapting the existing network using new-task training data, including adjusting its shared learn representation, thereby teaching using parameters of a pre-trained neural network as the starting model parameters for subsequent fine-tuning. )
Regarding claim 16, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim 1, wherein first optimization step and the second optimization step each comprise a gradient descent step (On page 3, Li teaches two optimization steps, with each optimization comprising a gradient-descent step.)
Regarding claim 17, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim 1, wherein the one or more data points of the primary domain include or are divided into a first set of data points for training the neural network model, a second set of data points for validating the neural network model and a third set of data points for testing the neural network model (On page 3, Li teaches minimizing a combined optimization objective comprising a first loss F(Θ) associated with meta-training domains and a second loss G(Θ) associated with virtual meta-test domains, wherein the losses are combined as F(Θ)+βG(Θ−αF(Θ)) to determine/update the model parameters. )
Regarding claim 19, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Li disclose:
The computer-implemented method of claim [[18]] 1, wherein the neural network model is trained by empirical risk minimization (ERM) (On page 3, Li teaches minimizing a combined optimization objective comprising a first loss F(Θ) associated with meta-training domains and a second loss G(Θ) associated with a virtual meta-test domain.)
Regarding claim 20, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Hoiem disclose:
A neural network trained in accordance with the method of claim [[18]]1 to perform the task in the first primary domain and the second primary domain (Fig.3 & Page 5 (Relationship to joint training), Hoiem discloses original/old-task training dataset new task dataset.)
With respect to claim 21, Li disclose:
Determining at least one set of auxiliary model parameters for the neural network model by simulating, using a processor, at least one first optimization step based on a set of current model parameters for the neural network model and at least one auxiliary domain, wherein the at least one auxiliary domain is associated with a primary domain comprising data points for training the neural network model (On page 3, Li teaches determining an updated set of model parameters by simulating an optimization step based on the current model parameters and training data associated with a domain. In particular, Li's meta-learning procedure performs a simulated optimization of the current model parameters using a meta-training domain to obtain updated model parameters, which are subsequently utilized in the meta-optimization procedure.)
Wherein the at least one set of auxiliary model parameters minimizes a loss associated with a respective auxiliary domain of the at least one auxiliary domain with respect to the current model parameters (On page 3 (Algorithm 1 & 2), Li teaches determining updated model parameters from current model parameters through a gradient-based optimization step and evaluating a loss G on a respective meta-test domain using the updated parameters. )
Determining a set of primary model parameters for the neural network model by performing, using the processor, a second optimization for the neural network model comprising minimizing a combined loss function, wherein the combined loss function combines: (On page 3, Li teaches a two-level optimization. First, it performs an inner/meta-train optimization from the current parameter, producing an intermediate parameter. Then Li constructs the overall meta-objective, where F represents the meta-training loss and G represents the meta-test loss evaluated after the simulated update.)
A first loss function associated with the set of current model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a first loss F(Θ) representing the meta-training loss using current parameters.)
A second loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a second loss G(Θ), using the updated/intermediate parameter and combining.)
Updating the neural network model with the set of primary model parameters (On page 3, Li teaches updating the model. The model is parameterized by Θ, so the gradient of Θ calculated with respect to this loss function is ∇Θ, and optimization will update the model as Θ=Θ−α∇Θ.)
With respect to claim 21, Li does not explicitly disclose:
A method for performing a task, the method comprising: performing, by a neural network model implemented by a processor and trained on the first primary domain in a first stage of training, the task in a first primary domain
Performing, by the trained neural network model trained by the processor on the first primary domain in the first stage of training and fine-tuned by the processor in a second stage of training on a second primary domain, the task in the first primary domain or the second primary domain
Wherein each of the first stage of training and a second stage of training comprises:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
Each of a first stage of training and a second stage of training
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data
However, it is known by Hoiem to disclose:
A method for performing a task, the method comprising: performing, by a neural network model implemented by a processor and trained on the first primary domain in a first stage of training, the task in a first primary domain ((Fig.3 & Page 5 (Relationship to joint training), Hoiem discloses original/old-task training dataset new task dataset.) On page 5, Hoiem teaches a convolutional neural network previously trained using original-task training data, thereby teaching a first stage of training in which the neural network is trained on data points of an original/first domain. Hoiem subsequently trains the network using only new-task data while the original-task training data are unavailable. )
Performing, by the trained neural network model trained by the processor on the first primary domain in the first stage of training and fine-tuned by the processor in a second stage of training on a second primary domain, the task in the first primary domain or the second primary domain (On page 4 (3 Learning without forgetting), Hoiem subsequently teaches subsequently training an already-trained convolutional neural network using data associated with a new task (i.e., a secondary primary domain), wherein the training data associated with the network's previously learned capabilities (i.e., data points of a first primary domain) are unavailable. Hoiem further teaches learning without forgetting and uses only the new-task data to train the network while preserving the previously learned capabilities, thereby teaching subsequent training in the second primary domain by accessing data points of the first primary domain.)
Each of the first stage of training and a second stage of training comprises:(Fig.3 & Page 5 (Relationship to joint training & efficiency comparison). Hoiem teaches by gradually adding new capabilities to an existing CNN when the training data for the existing capabilities are unavailable. Prior/original training that established CNN's existing capabilities.)
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain (On page 5, Hoiem teaches a convolutional neural network previously trained using original-task training data, thereby teaching a first stage of training in which the neural network is trained on data points of an original/first domain. Hoiem subsequently trains the network using only new-task data while the original-task training data are unavailable. )
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain (On page 4 (3 Learning without forgetting), Hoiem subsequently teaches subsequently training an already-trained convolutional neural network using data associated with a new task (i.e., a secondary primary domain), wherein the training data associated with the network's previously learned capabilities (i.e., data points of a first primary domain) are unavailable. Hoiem further teaches learning without forgetting and uses only the new-task data to train the network while preserving the previously learned capabilities, thereby teaching subsequent training in the second primary domain by accessing data points of the first primary domain.)
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task (The examiner selects: classification task: On page 2 (1–2 paragraphs), Hoiem teaches updating an existing convolutional neural network using new-task training data while preserving previously learned capabilities, wherein the resulting updated convolutional neural network performs visual classification tasks associated with the new and previously learned capabilities, thereby teaching an updates neural network model used to perform as least a classification task.)
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data (The examiner selects: image. On page 3 and Fig. 2, Hoiem teaches that data points of the training domains comprise image data, as the disclosed convolutional neural network is trained and evaluated using image datasets for computer-vision/classification tasks.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to train multiple source domains as taught by Li, while using new-task training data while attempting to preserve the network's existing capabilities as taught by Hoiem. The motivation for doing so would have been to improve training domain performance should also improve testing domain performance (See Page 1 of the Abstract of Li.)
With respect to claim 21, Li in view of Hoiem does not explicitly disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
However, it is known by Albright to disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation (In paragraphs [0033-0034], Albright teaches generating augmented training data by modifying images of an initial training dataset, wherein one or more transformations are performed on the initial training images to generate modified training images. The disclosed transformations include scaling, rotation, translation, brightness modification, lighting modification, and perspective modification. Thus, the Albright reference teaches or at least suggests selecting from a plurality of available transformations and manipulating data points of an original training domain using the selected transformation to generate modified training data.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, with one or more transformations on the portion of the image as taught by Albright. The motivation for doing so would have been to replace fine-tuning with similar old and new task datasets for improved new task performance (See (Abstract) of Hoiem.)
Regarding claim 22, Li in view Hoiem and Albright disclose the elements of claim 1. In addition, Hoiem disclose:
The method of claim 21, wherein the combined loss function further combines a third loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the at least one auxiliary domain (On page 3, Li teaches minimizing a combined optimization objective comprising a first loss F(Θ) associated with meta-training domains and a second loss G(Θ) associated with virtual meta-test domains, wherein the losses are combined as F(Θ)+βG(Θ−αF(Θ)) to determine/update the model parameters.)
With respect to claim 23, Li disclose:
An apparatus for training a neural network model comprising: a non-transitory computer-readable medium having executable instructions stored thereon for causing a processor and a memory to perform a method for sequentially learning a plurality of domains associated with a task, the plurality of domains comprising a < multiple source domains>, the method comprising (On Pages 2–3 (Meta-Learning Domain Generalization), Li teaches training a machine-learning model using a plurality of domains associated with a classification task. In particular, Li discloses a domain-generalization framework in which training data are obtained from multiple source domains and the sources are divided into meta-train and meta-test domains to simulate domain shifts during training, thereby training the model to generalize across different domains.)
Determining at least one set of auxiliary model parameters for the neural network model by simulating, using a processor, at least one first optimization step based on a set of current model parameters for the neural network model and at least one auxiliary domain, wherein the at least one auxiliary domain is associated with a primary domain comprising data points for training the neural network model (On page 3, Li teaches determining an updated set of model parameters by simulating an optimization step based on the current model parameters and training data associated with a domain. In particular, Li's meta-learning procedure performs a simulated optimization of the current model parameters using a meta-training domain to obtain updated model parameters, which are subsequently utilized in the meta-optimization procedure.)
Wherein the at least one set of auxiliary model parameters minimizes a loss associated with a respective auxiliary domain of the at least one auxiliary domain with respect to the current model parameters (On page 3 (Algorithm 1 & 2), Li teaches determining updated model parameters from current model parameters through a gradient-based optimization step and evaluating a loss G on a respective meta-test domain using the updated parameters. )
Determining a set of primary model parameters for the neural network model by performing, using the processor, a second optimization for the neural network model comprising minimizing a combined loss function, wherein the combined loss function combines: (On page 3, Li teaches a two-level optimization. First, it performs an inner/meta-train optimization from the current parameter, producing an intermediate parameter. Then Li constructs the overall meta-objective, where F represents the meta-training loss and G represents the meta-test loss evaluated after the simulated update.)
A first loss function associated with the set of current model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a first loss F(Θ) representing the meta-training loss using current parameters.)
A second loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a second loss G(Θ), using the updated/intermediate parameter and combining.)
Updating, using the processor, the neural network model with the set of primary model parameters (On page 3, Li teaches updating the model. The model is parameterized by Θ, so the gradient of Θ calculated with respect to this loss function is ∇Θ, and optimization will update the model as Θ=Θ−α∇Θ.)
With respect to claim 23, Li does not explicitly disclose:
<A first primary domain and a second primary domain>
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
Each of a first stage of training and a second stage of training
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data
However, it is known by Hoiem to disclose:
<A first primary domain and a second primary domain> known as multiple source domains (Fig.3 & Page 5 (Relationship to joint training), Hoiem discloses original/old-task training dataset new task dataset.)
Each of a first stage of training and a second stage of training (Fig.3 & Page 5 (Relationship to joint training & efficiency comparison). Hoiem teaches by gradually adding new capabilities to an existing CNN when the training data for the existing capabilities are unavailable. Prior/original training that established CNN's existing capabilities.)
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain (On page 5, Hoiem teaches a convolutional neural network previously trained using original-task training data, thereby teaching a first stage of training in which the neural network is trained on data points of an original/first domain. Hoiem subsequently trains the network using only new-task data while the original-task training data are unavailable. )
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain (On page 4 (3 Learning without forgetting), Hoiem subsequently teaches subsequently training an already-trained convolutional neural network using data associated with a new task (i.e., a secondary primary domain), wherein the training data associated with the network's previously learned capabilities (i.e., data points of a first primary domain) are unavailable. Hoiem further teaches learning without forgetting and uses only the new-task data to train the network while preserving the previously learned capabilities, thereby teaching subsequent training in the second primary domain by accessing data points of the first primary domain.)
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task (The examiner selects: classification task: On page 2 (1–2 paragraphs), Hoiem teaches updating an existing convolutional neural network using new-task training data while preserving previously learned capabilities, wherein the resulting updated convolutional neural network performs visual classification tasks associated with the new and previously learned capabilities, thereby teaching an updates neural network model used to perform as least a classification task.)
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data (The examiner selects: image. On page 3 and Fig. 2, Hoiem teaches that data points of the training domains comprise image data, as the disclosed convolutional neural network is trained and evaluated using image datasets for computer-vision/classification tasks.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to train multiple source domains as taught by Li, while using new-task training data while attempting to preserve the network's existing capabilities as taught by Hoiem. The motivation for doing so would have been to improve training domain performance should also improve testing domain performance (See Page 1 of the Abstract of Li.)
With respect to claim 23, Li in view of Hoiem does not explicitly disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
However, it is known by Albright to disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation (In paragraphs [0033-0034], Albright teaches generating augmented training data by modifying images of an initial training dataset, wherein one or more transformations are performed on the initial training images to generate modified training images. The disclosed transformations include scaling, rotation, translation, brightness modification, lighting modification, and perspective modification. Thus, the Albright reference teaches or at least suggests selecting from a plurality of available transformations and manipulating data points of an original training domain using the selected transformation to generate modified training data.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, with one or more transformations on the portion of the image as taught by Albright. The motivation for doing so would have been to replace fine-tuning with similar old and new task datasets for improved new task performance (See (Abstract) of Hoiem.)
With respect to claim 24, Li disclose:
A system for training a neural network model for sequentially learning a plurality of domains associated with a task, the plurality of domains comprising a first primary domain and a second primary domain comprising: a processor; a memory; and computer-executable instructions stored on a non-transitory computer-readable medium for causing the processor to perform a method comprising (On Pages 2–3 (Meta-Learning Domain Generalization), Li teaches training a machine-learning model using a plurality of domains associated with a classification task. In particular, Li discloses a domain-generalization framework in which training data are obtained from multiple source domains and the sources are divided into meta-train and meta-test domains to simulate domain shifts during training, thereby training the model to generalize across different domains. (Li is directed to computer-implemented neural-network training techniques involving gradient computations, loss evaluation, and parameter updates. A person of ordinary skill in the art would have understood these operations to be performed by one or more processors executing instructions stored in memory. ))
Determining at least one set of auxiliary model parameters for the neural network model by simulating, using a processor, at least one first optimization step based on a set of current model parameters for the neural network model and at least one auxiliary domain, wherein the at least one auxiliary domain is associated with a primary domain comprising data points for training the neural network model (On page 3, Li teaches determining an updated set of model parameters by simulating an optimization step based on the current model parameters and training data associated with a domain. In particular, Li's meta-learning procedure performs a simulated optimization of the current model parameters using a meta-training domain to obtain updated model parameters, which are subsequently utilized in the meta-optimization procedure.)
Wherein the at least one set of auxiliary model parameters minimizes a loss associated with a respective auxiliary domain of the at least one auxiliary domain with respect to the current model parameters (On page 3 (Algorithm 1 & 2), Li teaches determining updated model parameters from current model parameters through a gradient-based optimization step and evaluating a loss G on a respective meta-test domain using the updated parameters. )
Determining a set of primary model parameters for the neural network model by performing, using the processor, a second optimization for the neural network model comprising minimizing a combined loss function, wherein the combined loss function combines: (On page 3, Li teaches a two-level optimization. First, it performs an inner/meta-train optimization from the current parameter, producing an intermediate parameter. Then Li constructs the overall meta-objective, where F represents the meta-training loss and G represents the meta-test loss evaluated after the simulated update.)
A first loss function associated with the set of current model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a first loss F(Θ) representing the meta-training loss using current parameters.)
A second loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the primary domain (On page 3, Li teaches a second loss G(Θ), using the updated/intermediate parameter and combining.)
Updating the neural network model with the set of primary model parameters (On page 3, Li teaches updating the model. The model is parameterized by Θ, so the gradient of Θ calculated with respect to this loss function is ∇Θ, and optimization will update the model as Θ=Θ−α∇Θ.)
With respect to claim 24, Li does not explicitly disclose:
<A first primary domain and a second primary domain>
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
Each of a first stage of training and a second stage of training
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data
However, it is known by Hoiem to disclose:
<A first primary domain and a second primary domain> known as multiple source domains (Fig.3 & Page 5 (Relationship to joint training), Hoiem discloses original/old-task training dataset new task dataset.)
Each of a first stage of training and a second stage of training (Fig.3 & Page 5 (Relationship to joint training & efficiency comparison). Hoiem teaches by gradually adding new capabilities to an existing CNN when the training data for the existing capabilities are unavailable. Prior/original training that established CNN's existing capabilities.)
Wherein the first stage of training trains the neural network model on the data points of the primary domain being the first primary domain (On page 5, Hoiem teaches a convolutional neural network previously trained using original-task training data, thereby teaching a first stage of training in which the neural network is trained on data points of an original/first domain. Hoiem subsequently trains the network using only new-task data while the original-task training data are unavailable. )
Wherein the second stage of training subsequently trains the trained neural network model on the data points of the primary domain being the second primary domain without accessing data points of the first primary domain (On page 4 (3 Learning without forgetting), Hoiem subsequently teaches subsequently training an already-trained convolutional neural network using data associated with a new task (i.e., a secondary primary domain), wherein the training data associated with the network's previously learned capabilities (i.e., data points of a first primary domain) are unavailable. Hoiem further teaches learning without forgetting and uses only the new-task data to train the network while preserving the previously learned capabilities, thereby teaching subsequent training in the second primary domain by accessing data points of the first primary domain.)
Wherein the updated neural network model is used to perform one or more of a recognition task, a classification task, an autonomous movement task, or a natural language processing task (The examiner selects: classification task: On page 2 (1–2 paragraphs), Hoiem teaches updating an existing convolutional neural network using new-task training data while preserving previously learned capabilities, wherein the resulting updated convolutional neural network performs visual classification tasks associated with the new and previously learned capabilities, thereby teaching an updates neural network model used to perform as least a classification task.)
Wherein the data points of the primary domain comprise one or more of text, voice, image, or sensory data (The examiner selects: image. On page 3 and Fig. 2, Hoiem teaches that data points of the training domains comprise image data, as the disclosed convolutional neural network is trained and evaluated using image datasets for computer-vision/classification tasks.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to train multiple source domains as taught by Li, while using new-task training data while attempting to preserve the network's existing capabilities as taught by Hoiem. The motivation for doing so would have been to improve training domain performance should also improve testing domain performance (See Page 1 of the Abstract of Li.)
With respect to claim 24, Li in view of Hoiem does not explicitly disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation
However, it is known by Albright to disclose:
Wherein the at least one auxiliary domain comprises a plurality of data points modified from the primary domain via data manipulation, wherein the data manipulation comprises (a) selecting a transformation from a set of transformations, and (b) manipulating one or more data points of the primary domain using the selected transformation (In paragraphs [0033-0034], Albright teaches generating augmented training data by modifying images of an initial training dataset, wherein one or more transformations are performed on the initial training images to generate modified training images. The disclosed transformations include scaling, rotation, translation, brightness modification, lighting modification, and perspective modification. Thus, the Albright reference teaches or at least suggests selecting from a plurality of available transformations and manipulating data points of an original training domain using the selected transformation to generate modified training data.)
Li, in view of Hoiem, are analogous pieces of art because both references concern training/adapting neural-network models so that learned model parameters perform appropriately when the model encounters data/tasks that differ from the original training setting. Accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, with one or more transformations on the portion of the image as taught by Albright. The motivation for doing so would have been to replace fine-tuning with similar old and new task datasets for improved new task performance (See (Abstract) of Hoiem.)
Regarding claim 25, Li in view Hoiem and Albright disclose the elements of claim 23. In addition, Li disclose:
The apparatus of claim 23, wherein the combined loss function further combines a third loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the at least one auxiliary domain (On page 3, Li teaches algorithm divides the available domains into meta-train domains and meta-test domains. It first computes the meta-training loss F, takes a gradient step to obtain updates/auxiliary parameters, and then evaluates those updated parameters on the meta-test domains using a separate loss G. Li expressly identifies F as the loss from the aggregated meta-test domains.)
Regarding claim 26, Li in view Hoiem and Albright disclose the elements of claim 23. In addition, Li disclose:
The system of claim 24, wherein the combined loss function further combines a third loss function associated with the at least one set of auxiliary model parameters for the neural network model and associated with the at least one auxiliary domain (On page 3, Li teaches algorithm divides the available domains into meta-train domains and meta-test domains. It first computes the meta-training loss F, takes a gradient step to obtain updates/auxiliary parameters, and then evaluates those updated parameters on the meta-test domains using a separate loss G. Li expressly identifies F as the loss from the aggregated meta-test domain.
Claim 6 is rejected under 35 U.S.C 103 as being unpatentable over Li in view of Hoiem, Albright and further in view of Kumar et al. (US Pub No.: 20170300783 A1), hereinafter referred to as Kumar.
Regarding claim 6, Li in view of Hoiem and Albright disclose element of claim 1. Li in view of Hoiem and Albright do not explicitly disclose:
The computer-implemented method of claim [[2]] 1, wherein the data manipulation comprises at least one image transformation
Wherein the at least one image transformation comprises at least one of a photometric and a geometric transformation.
However, Kumar disclose the limitation:
The computer-implemented method of claim [[2]] 1, wherein the data manipulation comprises at least one image transformation (In paragraph [0028], Kumar discloses that a transformation operation, as shown in block 34, can be implemented with respect to the source images.)
Wherein the at least one image transformation comprises at least one of a photometric and a geometric transformation (In paragraph [0051], Kumar discloses the aforementioned characteristics that can include image characteristics such as a geometric transformation and a photometric transformation.)
Accordingly, it would have been obvious to a person having ordinary skill in the art
before the effective filing date of the claimed invention having the teachings of Li in view of Hoiem and Albright before them, to include Kumar, with generic geometric and photometric transformations, which make use of only source domain data as taught by Kumar. The motivation for doing so would have been to improve image classification accuracy in a new domain (See [0009] of Kumar.)
Claim 15 is rejected under 35 U.S.C 103 as being unpatentable over Li in view of Hoiem, Albright and further in view of WANG et al. (US Pub No.: 20170228645 A1), hereinafter referred to as WANG.
Regarding claim 15, Li in view of Hoiem and Albright disclose element of claim 1. Li in view of Hoiem and Albright do not explicitly disclose:
The computer-implemented method of claim 1,wherein, in each of the first stage of training and the second stage of training, the steps of determining at least one set of auxiliary model parameters for the neural network model, determining a set of primary model parameters for the neural network model, and updating the neural network model are repeated until at least one of a gradient descent step size for the second optimization for the neural network model is below a threshold and a maximum number of gradient descent steps is reached.
However, WANG disclose the limitation (In paragraph [0048], WANG discloses that teaches a convolutional neural network using inconsistent stochastic gradient descent (IGD) and repeatedly training/updating it. Critically, it teaches continuing training until the determined loss falls below a predetermined threshold or a predetermined number of iterations is reached.)
Accordingly, it would have been obvious to a person having ordinary skill in the art
before the effective filing date of the claimed invention having the teachings of Li in view of Hoiem and Albright before them, to include WANG, with training a convolutional neural network using an inconsistent stochastic gradient descent (ISGD) algorithm as taught by WANG. The motivation for doing so would have been to improve method for training convolutional neural networks employing inconsistent stochastic gradient descent (See [0002] of WANG).
Response to Arguments
The applicant's arguments filed 05/19/2026 have been fully considered, but in part are not persuasive.
Pertaining to Rejection under 101
Rejections for claims 1, 3, 5-9, 11, 15-17, 19-26 are withdrawn under 35 USC § 101.
Pertaining to Rejection under 103
Applicant’s arguments in regard to the examiner’s rejections under 35 USC 103 are moot in view of the new grounds of rejection
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to EVEL HONORE whose telephone number is (703)756-1179. The examiner can normally be reached Monday-Friday 8 a.m. -5:30 p.m.
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, Mariela D Reyes can be reached at (571) 270-1006. 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.
EVEL HONORE
Examiner
Art Unit 2142
/Mariela Reyes/Supervisory Patent Examiner, Art Unit 2142
Prosecution Insights