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 in response to the amendment filed on 06/10/2025. Claims 1-4, 6-13, 15-20 are pending in the case. This action is Final.
Applicant Response
3. In Applicant’s response dated 06/10/2025, Applicant amended Claims 1, 6, 10, 14 and 20, cancelled claim 5 and 14 and argued against all objections and rejections previously set forth in the Office Action dated 03/11/2025.
Response to Amendment/Remarks
4. Claim Rejections - 35 USC § 101
The Applicant submits that claims 1-4, 6-13 and 15-20 stand rejected under 35 U.S.C. 101 on the basis that the claims are allegedly directed to an abstract idea without significantly more. The Applicant respectfully disagrees that the above elements can be characterized as being a recitation of a mental process (i.e., evaluation) and math. These are computer steps to improve computer learning processes to improve computer operation. They are statutory. Additionally, MPEP 2106.04(a)(1) cautions that: "Some claims are not directed to an abstract idea because they do not recite an abstract idea, although it may be apparent that at some level they are based on or involve an abstract idea. Because these claims do not recite an abstract idea (or other judicial exception), they are eligible at Step 2A Prong One (Pathway B)." The Applicant submits that there is no abstract idea recited in claim 1. The Applicant submits that claim 1 should be found to be directed to patent eligible subject matter at least under Step 2A Prong One.
Examiner respectfully disagrees
As the office action below, the claim requires a list of mathematical concepts as the table below shows .
training the encoder and the predictor over multiple interaction
Abstract idea (data processing , math)
inputting a set of tokens … from a multi-domain the encoder
Abstract idea (math, data manipulation)
obtaining, from the encoder … to encode domain- related information
Abstract idea (Mathematical concept)
inputting the unique embedding vector to an adaptor network to generate a set of domain probabilities
Probability computation, math, abstract
computing a domain mixing loss that represents an error
Loss computation, math formula
extracting a weights matrix from the adaptor network
Mathematical calculation
computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors
Mathematical calculation
inputting at least the domain mixing embedding vector containing domain-related information … predicted output based on at least the domain-related information;
Mathematical calculation
computing an output prediction loss
Mathematical calculation
computing a final loss using the domain mixing loss and the output prediction loss
Mathematical calculation
updating values of parameters of the encoder and the predictor of the neural network model
Gradient update = math concept
the training is performed until a convergence condition is satisfied
Convergence criteria= math
storing the updated values of the parameters of the encoder
Generic computer function
executing the neural network model, using the learned values of the parameters of the neural network model,
Abstract, using model for task
As shown above the claim the claim is directed to a judicial exception (mathematical idea mathematical concepts, mental processes. The claim does not recite a specific improvement to computer functioning or a practical technological field. The claim shows and describes how to train a model on a broad task on data across domains.
Under broadest reasonable interpretation, applying an auto-encoder model is adding the words “apply it” (or an equivalent) with the judicial exception, or merely using an auto-encoder model as a tool to perform an abstract idea (e.g., identify error) (see MPEP 2106.05). Updating the model to adapt is describing training a machine learning model which is understood as mere instructions to implement an abstract idea on a computer (see MPEP 2106.05(f).).) Applying an auto-encoder model is adding the words “apply it” (or an equivalent) with the judicial exception, or merely using an auto-encoder model as a tool to perform an abstract idea (see MPEP 2106.05). Training a machine learning model is understood as mere instructions to implement an abstract idea on a computer (see MPEP 2106.05(f).).)
It is not clear how the recited claim language discloses a specific solution and appears to be describing generally applying a model and adapting the model to changes. The cited specification paragraph recites features not incorporated into the claims and it is not clear which claim limitations are being associated with the cited specification paragraph. Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims.
Therefore, these limitations, taken alone or in combination, do not integrate the abstract idea into a practical application or recite significantly more that the abstract idea. Thus, all of the dependent claims are also not patent eligible.
2. Claim Rejection 35 U.S.C. § 103
Applicant’s prior art arguments see pages 16-20, filed on 09/15/2025, with respect to claims 1, 10, and 20 have been fully considered but they are not persuasive.
Applicants’ 35 U.S.C. § 103 arguments have been fully considered but they are not persuasive. Specifically, the application asserts that the cited references teachings are not sufficient to render the claims prima facie obvious. Examiner respectfully disagrees. In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
Further, the applicant asserts that the cited reference of Chen fails to disclose “Thus, Chen fails to disclose training a neural network model including "inputting a set of tokens from a data sample sampled from a multi-domain training dataset that includes data from multiple single-domain datasets to the encoder of the neural network model, the set of tokens including a unique token and other tokens" and "obtaining, from the encoder, a set of embedding vectors including a unique embedding vector encoded from the unique token and other embedding vectors encoded from the other tokens, wherein the encoder is trained over the plurality of training iterations to encode domain-related information in the unique embedding vector" as recited in the Applicant's claim 1.”
Examiner respectfully disagrees. As the office action shows, Chen does teach a method “inputting a set of tokens from a data sample sampled from a multi-domain training dataset that includes data from multiple single-domain datasets to the encoder of the neural network model, the set of tokens including a unique token and other tokens (see CHEN: Fig.1, [0033], “An input to the lexicon encoder can be sequence of tokens of length m, X= {x.sub.1, . . ., x.sub.m}. Specific tokens (unique tokens) can be used to delineate the beginning of each sequence, and to separate individual sentences in a given sequence)…[0053], “, input embedding vectors 306 and context embedding vectors 308 can provide representations of words, sentences, phrases, or, more generally, n-grams or other tokens, that can be readily extended to new tasks by the domain adaptation techniques described previously.”)
Further the application asserts that the cited references of "Chen and Bui does not teach the method wherein: extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains; computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain". Son is relied upon to disclose these claim elements.
Examiner respectfully disagrees. In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986). As the office action below shows, Sun teaches “extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains(see SUN: Fig.3, [0085], “In operation 320, the training apparatus may extract an embedding vector (a feature vector, for example) corresponding to the training data received in operation 310, by encoding the training data (using an encoder, for example).”; computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain (see SUN: Fig.4, [0109], “The training apparatus 400 may train the ANN to minimize the reconstruction loss together with the above-described target loss. For example, the training apparatus 400 may train the ANN to minimize a sum (for example, a weighted sum) of a reconstruction loss and a target loss for a full set of training data.”)
Therefore, Examiner respectfully asserts that the cited art sufficiently teaches the limitations recited in the presented claims. Therefore, the given 35 U.S.C. 103 rejection has been remains sustained for 1-4, 6-13 and 15-20.
Claim Rejections - 35 USC § 101
4. 35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-4, 6-13 and 15-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1 Analysis: Is the claim to a process, machine, manufacture or composition of matter? See MPEP § 2106.03.
Claims 1-4, 6-9 are directed to a method claim, claims 10-13 and 15-19 is directed to a computer system claim and claim 20 is directed to a non-transitory computer readable storage medium. Thus, each of the claims falls within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
Step 2A Prong 1 Analysis: Does the claim recite an abstract idea, law of nature, or natural phenomenon? See MPEP § 2106.04(II)(A)(1).
Claim 1 recites :
“inputting the unique embedding vector to an adaptor network to generate a set of domain probabilities representing a likelihood that the unique embedding vector belongs to each domain of a set of multiple domains (This step for performing imputations is practically implementable math and is understood to be a recitation of a mental process or mathematical concept (i.e., evaluation) and math.);
computing a domain mixing loss that represents an error between the set of domain probabilities and a ground-truth domain of the data sample” (This step for calculating an errors is implementable in the human mind mathematical calculation and is understood to be a recitation of a mathematical process (i.e., math/evaluation)
extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains (This step for performing imputations is practically implementable in the mathematical concept and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain (This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
“inputting at least a domain mixing embedding vector containing domain-related information, determined from the unique embedding vector, to the predictor of the neural network model, to generate a predicted output based on at least the domain-related information This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
“computing an output prediction loss that represents an error between the predicted output and a ground-truth label of the data sample” This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
“computing a final loss using the domain mixing loss and the output prediction loss” This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.) and
“updating values of parameters of the encoder and the predictor of the neural network model and also updating values of parameters of the adaptor network, using the computed final loss”( This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
wherein the training is performed until a convergence condition is satisfied This step for performing mathematical operation/calculations and data processing techniques and is understood to be a recitation of a mental process (i.e., evaluation) and math.);
Step 2A prong 2 Analysis: Does the claim recite additional elements? Do those additional elements, individually and in combination, integrate the judicial exception into a practical application?
The claim recites additional element :
“training the encoder and the predictor of the neural network model over a plurality of training iterations to perform a task”,
“inputting a set of tokens from a data sample sampled from a multi-domain training dataset that includes data from multiple single-domain datasets to the encoder of the neural network model, the set of tokens including a unique token and other tokens”, “obtaining, from the encoder, a set of embedding vectors including a unique embedding vector encoded from the unique token and other embedding vectors encoded from the other tokens”, “inputting the unique embedding vector to an adaptor network to generate a set of domain probabilities representing a likelihood that the unique embedding vector belongs to each domain of a set of multiple domains.”, “storing the updated values of the parameters of the encoder and the predictor of the neural network model as learned values of the parameters of the neural network model”, and “executing the neural network model, using the learned values of the parameters of the neural network model, to perform the task on data across the set of multiple domains”. These additional limitations fail to integrate the abstract idea into a practical application. The limitations, recited at a high level of generality, only amount to “apply it” using a generic computer components (MPEP 2106.05(f)).
Step 2B Analysis: The claim does not include additional elements that are sufficient to amount
to significantly more than the judicial exception.
The additional elements as disclosed above in combination of the abstract idea are not sufficient to amount to significantly more than the judicial exception as they are generic computer functions in combination with limitations that are generally linking the use of the judicial exception to a particular technological environment or field of use that are implemented to perform the disclosed abstract idea above. The additional elements, alone and in combination, fail to integrate the abstract idea into a practical application. Thus, the claims are not patent eligible.
Regarding independent Claim 10: see the rejection of claim 1 above. Same rationale applies.
2A Prong 2 & 2B: The claim recites another additional element “A computing system for training a neural network model having an encoder and a predictor, the computing system comprising a processing unit and a memory storing instructions which, when executed by the processing unit, cause the computing system:” (mere instructions to apply the exception using a generic computer component-see MPEP 2106.05(f))
Regarding independent Claim 20: see the rejection of claim 1 above. Same rationale applies.
2A Prong 2 & 2B: The claim recites another additional element “A non-transitory computer readable medium having instructions encoded thereon, wherein the instructions, when executed by a processing unit of a computing system, cause the computing system to ... (mere instructions to apply the exception using a generic computer component-see MPEP 2106.05(f))
The dependent claims respectively recite a judicial exception in limitations of: “wherein the predictor is a decoder, and wherein the other embedding vectors are also inputted to the decoder to generate the predicted output.” (claims 2,11), “wherein the predictor is a classifier, and only the domain mixing embedding vector is inputted to the classifier to generate the predicted output.” (claims 3,12), “wherein the domain mixing embedding vector is the unique embedding vector.” (claims 4,13), “inputting the set of tokens to each of a plurality of teacher models, to generate a respective set of logits from each teacher model, each teacher model being pre-trained in a respective single domain of the set of domains; and computing at least one of a distillation loss or a contrastive loss using at least one set of logits from one teacher model and a set of logits generated by the predictor; wherein the at least one of the distillation loss or the contrastive loss is further included in computing the final los” (claims 6,15), “wherein the distillation loss is computed using the set of logits generated by the predictor and the set of logits generated by an in-domain teacher model, the in-domain teacher model being the teacher model that is pre- trained in the domain corresponding to the ground-truth domain of the data sample.” (claims 7,16), “wherein the distillation loss is computed using the set of logits generated by the predictor and a weighted aggregation of the sets of logits from the plurality of teacher models, wherein each set of logit generated by a respective teacher model is weighted by the domain probability corresponding to the domain of the respective teacher model.” (claims 8,17) “wherein both the distillation loss and the contrastive loss is computed, and both the distillation loss and the contrastive loss are further included in computing the final loss.” (claims 9,18) “, wherein the computing system provides a cloud-based service for training the neural network model.” (claim 19). These additional limitations (in claims 1-9 and 11-19) also constitute concepts performed in the human mind which fall within the “Mental Processes” groupings of abstract ideas.
This judicial exception is not integrated into a practical application. Additional elements
“computing system comprising a processing unit and a memory storing instructions which, when executed by the processing unit, (in claims 11-3 and 15-19) and “non-transitory computer readable medium comprising: computer program code” (in claim 20), all amount to no more than adding insignificant extra-solution activity/specifications related to data gathering, data input, or data transmittal. These additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea. The dependent claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of non-transitory computer readable medium comprising: computer program code are again insignificant extra-solution activity steps that cannot provide an inventive concept. All of these additional elements as generically claimed are considered well-understood, routine, and conventional.
Therefore, these limitations, taken alone or in combination, do not integrate the abstract idea into a practical application or recite significantly more that the abstract idea. Thus, all of the dependent claims are also not patent eligible.
Examiner Comments
4. In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
6. Claims 1-4, 10-12 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Chen (US 20200334520 A1: 2020-10-22) in view of Bui ( US 20210271822 A1, 2021-09-02) in further view of SUN (US 20220108136 A1: 2022-04-07 )in further view of Li (US 20220391768 A1, 2022-12-08) in further view of ZHAO (US 20220198339 A1, 2022-06-23)
Regarding independent Claim 1,
Chen teaches a method for training a transformer neural network model having an encoder and a predictor, the method comprising:
training the encoder and the predictor of the neural network model over a plurality of training iterations to perform a task (see CHEN: Fig.4-6,[0044], raining can proceed with the first training data set, which can be used to train single-sentence classification layer 310(1), … lexicon encoder 304(1), and transformer encoder 304(2). See also [0042], stating “in each epoch (iteration), a mini-batch but of labeled task-specific data is selected, and the multi-task machine learning model is updated according to the task-specific objective for that task t.”), wherein performing each training iteration comprises:
inputting a set of tokens from a data sample sampled from a multi-domain training dataset that includes data from multiple single-domain datasets to the encoder of the neural network model, the set of tokens including a unique token and other tokens (see CHEN: Fig.1, [0033], “An input to the lexicon encoder can be sequence of tokens of length m, X= {x.sub.1, . . ., x.sub.m}. Specific tokens (unique tokens) can be used to delineate the beginning of each sequence, and to separate individual sentences in a given sequence)…[0053], “, input embedding vectors 306 and context embedding vectors 308 can provide representations of words, sentences, phrases, or, more generally, n-grams or other tokens, that can be readily extended to new tasks by the domain adaptation techniques described previously.”)
obtaining, from the encoder, a set of embedding vectors including a unique embedding vector encoded from the unique token and other embedding vectors encoded from the other tokens (see CHEN: Fig.1, [0033], “The lexicon encoder can map X into a sequence of input embedding vectors, one for each token. In some implementations, the input embedding vectors are constructed by summing corresponding word, segment, and positional embeddings for each word.”), wherein the encoder is trained over the plurality of training iterations to encode domain-related information in the unique embedding vector (see CHEN: Fig.1, [0034], “the transformer encoder is a multilayer bidirectional transformer encoder that is configured to map the input embedding vectors 306 into the context embedding vectors. The context embedding vectors can be used as a shared representation of the input phrases or sentences across different tasks.”)
wherein the training is performed until a convergence condition is satisfied (see CHEN: Fig.1, [0026], “Method 200 continues at block 208, where the trained multi-task machine learning model is output. For example, the trained multi-task machine learning model can be finalized and deployed when a convergence condition is reached and/or available training data is exhausted.)”
executing the neural network model, using the learned values of the parameters of the neural network model, to perform the task on data across the set of multiple domains (see CHEN: Fig.10, [0083], “The model training module 1021 can output a trained, final multi-task machine learning model to server 1030. Model execution module 1031 can execute the final multi-task machine learning model in response to received inputs. For example, server application 1032 can call the model execution module with one or more input data items to be processed by the model. ”
As shown above, Chen discloses the system of training and employing multi-task machine learning models relating to multi-task neural networks for natural language processing. However, as discussed thereafter, the present concepts can be employed in neural networks that perform other types of tasks, such image processing, audio processing, radar processing, etc.
Chen does not teach or disclose the system wherein:
inputting the unique embedding vector to an adaptor network to generate a set of domain probabilities representing a likelihood that the unique embedding vector belongs to each domain of a set of multiple domains;
computing a domain mixing loss that represents an error between the set of domain probabilities and a ground-truth domain of the data sample;
extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains;
computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain;
inputting at least the domain mixing embedding vector containing domain-related information, determined from the unique embedding vector, to the predictor of the neural network model, to generate a predicted output based on at least the domain-related information;
computing an output prediction loss that represents an error between the predicted output and a ground-truth label of the data sample;
computing a final loss using the domain mixing loss and the output prediction loss; and
updating values of parameters of the encoder and the predictor of the neural network model and also updating values of parameters of the adaptor network, using the computed final loss;
storing the updated values of the parameters of the encoder and the predictor of the neural network model as learned values of the parameters of the neural network model; and
However, Bui teaches the method wherein:
inputting the unique embedding vector to an adaptor network to generate a set of domain probabilities representing a likelihood that the unique embedding vector belongs to each domain of a set of multiple domains (see Bui: Fig.1, [0058], “the control module 13 may calculate each entire representation vector and multi-task learning module 14 may feed the entire representation vector to a task-specific classifier (an adaptor network) to compute the probability distribution for the possible labels for each task and perform knowledge transfer between the first and second encoders using similarity between the MD task and the WSD resolving task.”)
computing a domain mixing loss that represents an error between the set of domain probabilities and a ground-truth domain of the data sample (see Bui: Fig.5, [0107], The first and second encoders 15 and 16 may perform the method as described in FIG. 4. A multi-task learning module 14 may perform knowledge transfer between the first and second encoders 15 and 16. The task-specific classifier 17 may compute a probability distribution for possible labels for each task. The multi-task learning module 14 may minimize the following loss function,”)’
Because both Chen and Bui are in the same/similar field of endeavor of Training a neural network model, accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to modify the teaching of Chen to include the system that “input the unique embedding vector to an adaptor network to generate a set of domain probabilities representing a likelihood that the unique embedding vector belongs to each domain of a set of domains” and computing a domain mixing loss as taught by Bui. One would have been motivated to make such a combination in order to provide users with efficient and robust Machine learning model by to handle diverse input data high accuracy using deep learning. (see Bui [0084])
Chen and Bui does not teach the method wherein:
extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains;
computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain;
inputting at least a domain mixing embedding vector containing domain-related information, determined from the unique embedding vector, to the predictor of the neural network model, to generate a predicted output based on at least the domain-related information;
computing an output prediction loss that represents an error between the predicted output and a ground-truth label of the data sample;
computing a final loss using the domain mixing loss and the output prediction loss; and
updating values of parameters of the encoder and the predictor of the neural network model and also updating values of parameters of the adaptor network, using the computed final loss;
storing the updated values of the parameters of the encoder and the predictor of the neural network model as learned values of the parameters of the neural network model; and
However, SON teaches the system wherein:
extracting a weights matrix from the adaptor network, each row of the weights matrix defining a respective a domain embedding vector representing each respective domain in the set of domains (see SON: Fig.3, [0085], “In operation 320, the training apparatus may extract an embedding vector (a feature vector, for example) corresponding to the training data received in operation 310, by encoding the training data (using an encoder, for example).”
computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors, each domain embedding vector being weighted by the respective domain probability for the respective domain (see SUN: Fig.4, [0109], “The training apparatus 400 may train the ANN to minimize the reconstruction loss together with the above-described target loss. For example, the training apparatus 400 may train the ANN to minimize a sum (for example, a weighted sum) of a reconstruction loss and a target loss for a full set of training data.”)
Because Chen Bui and SUN are in the same/similar field of endeavor of Training a neural network model, accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to modify the teaching of Chen to include the system that extracting a weights matrix from the adaptor network and computing a domain mixing embedding vector as a weighted sum of the domain embedding vectors as taught by Bui. One would have been motivated to make such a combination in order to provide users with efficient and robust Machine learning model by to handle diverse input data high accuracy using deep learning.
computing a domain mixing loss that represents an error between the set of domain probabilities and a ground-truth domain of the data sample;
inputting at least a domain mixing embedding vector containing domain-related information, determined from the unique embedding vector, to the predictor of the neural network model, to generate a predicted output based on at least the domain-related information;
computing an output prediction loss that represents an error between the predicted output and a ground-truth label of the data sample;
However, Li teaches the system wherein:
computing a domain mixing loss that represents an error between the set of domain probabilities and a ground-truth domain of the data sample (see Li: Fig.3, [0035], “The loss module 114 (domain mixing loss) may quantify an accuracy of the feature probability distribution generated for the new domain data 122 by considering absolute differences between the feature probability distribution and the ground truth data 124. Alternatively, or additionally, the loss module 114 may calculate a mean squared error of the feature probability distribution for new domain data 122 relative to the corresponding ground truth data 124. In this manner, the loss module 114 is configured to monitor the effectiveness of the domain transfer module 112 using any suitable type of loss function, such as likelihood loss, cross entropy loss, L1 loss, squared loss.”)
inputting at least a domain mixing embedding vector containing domain-related information, determined from the unique embedding vector, to the predictor of the neural network model, to generate a predicted output based on at least the domain-related information (see Li: Fig.2, [0039], “The domain adaptation system 104 receives the domain-specific model 118 and a new domain dataset 120 for use in generating the domain-agnostic model 106. As described herein, the domain-specific model 118 is representative of a machine learning model trained to generate predicted outputs, according to an original task or objective, from input data that is defined by at least one data type having one or more enumerated data values.”)
computing an output prediction loss that represents an error between the predicted output and a ground-truth label of the data sample (see Li: Fig.5 [0082], “The loss function 206 may be further refined based on one or more losses used to train the domain-specific model 118, such as one or more of the local loss 326 or global loss 330, as illustrated in FIG. 3 via the double-headed arrow connecting loss function 206 to the domain-specific model 118. The loss module 116 is configured to vary a schedule and/or ratio by which the loss function 206 is refined based on one or more of the local loss 326 or the global loss 330. The schedule and ratio by which the loss function 206 is refined based on the local loss 326 and/or the global loss 330 may be dependent on a task of interest for which the domain-specific model 118 is trained and/or the auxiliary task for which the domain-agnostic model 106 is trained to process different domain data.”);
updating values of parameters of the encoder and the predictor of the neural network model and also updating values of parameters of the adaptor network, using the computed final loss (see Li: Fig.5, [0083], “at least one parameter of the machine learning model is then updated using the loss function (block 512). The training module 116, for instance, tunes at least one parameter of the domain-agnostic model 106 based on the loss function 206. In some implementations, parameters of the domain-agnostic model 106 are initialized by the training module 116 based on parameters of the domain-specific model 118. In this manner, prior to updating by the training module 116 based on the loss function 206, the domain-agnostic model 106 may be representative of a copy of the domain-specific model 118.”)
storing the updated values of the parameters of the encoder and the predictor of the neural network model as learned values of the parameters of the neural network model (see Li: Fig.5, [0085], “The domain-agnostic model 106 may be output to local storage of the computing device implementing the domain adaptation system 104, such as computing device 102. Alternatively, or additionally, the domain-agnostic model 106 may be output to one or more storage locations that are remote from the computing device 102, such as to remote storage location 126, or to a different computing device, via network 128.”); and
Because Chen, Bui and Li are in the same/similar field of endeavor of Training a neural network model, accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to modify the teaching of Chen to include the system that compute loss in training neural network model and updating values of parameters of the neural network model storing the updated values of the parameters of the neural network model as taught by Li. One would have been motivated to make such a combination in order to provide users with efficient and accurate Machine learning model by enabling generation of a trained model configured to handle diverse input data without requiring the size and scope of training data otherwise necessitated by conventional approaches, thereby reducing an amount of computational and network resources used in training a model. (see Li [0024].
Chen, Bui, SUN, and Li does not teach the system wherein:
computing a final loss using the domain mixing loss and the output prediction loss.
However, ZHAO teaches the system wherein:
computing a final loss using the domain mixing loss and the output prediction loss (see ZHAO: Fig.2, [0057], “the training model 130 may be trained based on the first loss function 210 and the second loss function 220. That is, a total loss function of the training model 130 may be determined based on the first loss function 210 and the second loss function 220. For example, the total loss function of the training model 130 may be a sum of the two loss functions (i.e., the first loss function 210 and the second loss function 220). As another example, the two loss functions may be assigned weights, and the total loss function of the training model 130 may be determined based on the weights. In some embodiments, the weights of the two loss functions may be preset (e.g., by a user via a terminal device) to reflect the importance of the first processing unit 132 and the adversarial unit 133 during the training.”)
Because Chen, Bui, SUN, Li and ZHAO are in the same/similar field of endeavor of Training a neural network model, accordingly, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to modify the teaching of Chen to include the system that computing a final loss using the domain mixing loss and the output prediction loss as taught by ZHAO. One would have been motivated to make such a combination in order to provide users with efficient and accurate Machine learning model.
Regarding Claim 2,
Chen, Bui, SUN, Li and ZHAO teaches all the limitations of Claim 1. Chen further teaches the method wherein the predictor is a decoder, the other embedding vectors are also inputted to the decoder to generate the predicted output (see Chen: Fig.3, [0035], “The task-specific layers 310 (decoder) can include a single-sentence classification layer 310(1), a pairwise text similarity layer 310(2), and a pairwise text classification layer 310(3). Each task-specific layer can produce a corresponding task-specific output 312, e.g., a single-sentence classification output 312(1), a pairwise text similarity output 312(2), and a pairwise text classification output 312(3).”
Regarding Claim 3,
Chen, Bui, SUN, Li and ZHAO teaches all the limitations of Claim 1. Li further teaches the method wherein the predictor is a classifier, and only the domain mixing embedding vector is inputted to the classifier to generate the predicted output (see Li: Fig.3, [0057], “To generate the local context vectors 316, the domain-specific model 118 implements a local domain classifier 322, which is representative of a fully-convolutional neural network configured to output a domain prediction map 324 having a same size (e.g., width and height) as the input data 302. In some implementations the local domain classifier 322 is trained using local loss 326. The local loss 326 is representative of any suitable type of loss algorithm for aligning low-level features, such as a least-squares loss. The local context vector 316 is extracted from a middle layer of the local domain classifier 322, and includes information providing context for the local features 306 of the input data 302.”)
One would have been motivated to combine Chen and Li in order to provide users with efficient and robust Machine learning model by enabling enable generation of a trained model configured to handle diverse input data without requiring the size and scope of training data otherwise necessitated by conventional approaches, thereby reducing an amount of computational and network resources used in training a model. (see Li [0024])
Regarding Claim 4,
Chen, Bui, SUN, Li and ZHAO teaches all the limitations of Claim 1. Chen further teaches the method wherein the domain mixing embedding vector is the unique embedding vector (see Li: Fig.2, [0039], “The domain adaptation system 104 receives the domain-specific model 118 and a new domain dataset 120 for use in generating the domain-agnostic model 106. As described herein, the domain-specific model 118 is representative of a machine learning model trained to generate predicted outputs, according to an original task or objective, from input data that is defined by at least one data type having one or more enumerated data values.”)
One would have been motivated to combine Chen and Li in order to provide users with efficient and robust Machine learning model by enabling enable generation of a trained model configured to handle diverse input data without requiring the size and scope of training data otherwise necessitated by conventional approaches, thereby reducing an amount of computational and network resources used in training a model. (see Li [0024])
Regarding independent Claim 10,
Claim 10 is directed to a computer system and has the same/similar claim limitation/s as Claim 1 and is rejected under the same rationale.
Regarding Claim 11,
Claim 11 is directed to a computing system claim and has the same/similar claim limitation as Claim 2 and is rejected under the same rationale.
Regarding Claim 12,
Claim 12 is directed to a computing system claim and has the same/similar claim limitation as Claim 3 and is rejected under the same rationale.
Regarding Claim 13,
Claim 13 is directed to a computing system claim and has the same/similar claim limitation as Claim 4 and is rejected under the same rationale.
Regarding Claim 19,
Chen, Bui, SUN, Li and ZHAO teaches all the limitations of Claim 10. Chen further teaches the method wherein the computing system provides a cloud-based service for training the neural network model (see Chen: Fig.8, [0058], “illustrates an example method 800, consistent with the present concepts. Method 800 can be implemented on many different types of devices, e.g., by one or more cloud servers, by a client device such as a laptop, tablet, or smartphone, or by combinations of one or more servers, client devices, etc.”)
Regarding independent Claim 20,
Claim 20 is directed to a non-transitory computer readable medium and has the same/similar claim limitation as Claim 1 and is rejected under the same rationale.
Claims 6-9 and 15-18 are rejected under 35 U.S.C. 103 as being unpatentable over Chen, Bui, SUN, Li and ZHAO as applied to claims 1-4, 10-13 and 19-20 as shown above and in further view of Yan (DOCUMENT ID US 11200497 B1, DATE PUBLISHED 2021-12-14)
Regarding Claim 6,
As shown above Chen and Li teach all the limitations of Claim 1. Chen and Li does not teach the method comprising:
inputting the set of tokens to each of a plurality of teacher models;
computing at least one of a distillation loss or a contrastive loss using at least one set of logits from one teacher model and a set of logits generated by the predictor wherein the at least one of the distillation loss or the contrastive loss is further included in computing the final loss.
However, Yan teaches the method comprising:
inputting the set of tokens to each of a plurality of teacher models (see Yan: Fig.4A, Col.10, Line 50-53, “the teacher network 420 may include an input layer, an embedding layer, one or more self-attention layers, one or more feedforward layers, one output layer, and other layers.”), to generate a respective set of logits from each teacher model (see Yan: Fig.3, Col.9, Line 35-40, “soft cross-entropy loss between the logits of the student network and the teacher network (temp is the temperature parameter in the context of knowledge distillation, z.sup.T is the prediction logits of the teacher network, and z.sup.S is the prediction logits of the student network).”), each teacher model being pre-trained in a respective single domain of the set of domains (see Yan: Fig.3, Col.10, Line 34-42, “ The teacher network is a dense (i.e., large) neural network possessing knowledge learned from a task-specific domain as well as knowledge learned from other domains that are related to the task-specific domain. For example, the task-specific domain may refer to drafting sci-fi literature, and the “other domains” may refer to drafting fictional literature.”) and
computing at least one of a distillation loss or a contrastive loss using at least one set of logits from one teacher model and a set of logits generated by the predictor (see Yan: Fig.3, Col.8, Line 66-67 and Col.9 Line 1-12, “a distillation loss function may be defined to quantify the “difference between the output of the student network and the output of the one or more layers of the teacher network.” For example, the distillation loss function may include a plurality of loss functions corresponding to one or more layers of the teacher network. In some embodiments, the one or more layers of the teacher network may include: an embedding layer, a self-attention layer, a feedforward layer, an output layer, another layer, or any combination thereof. In some embodiments, the “one or more layers of the teach network” must include the output layer for knowledge distillation (i.e., knowledge distillation must be performed at the output layer).” wherein the at least one of the distillation loss or the contrastive loss is further included in computing the final loss (see Yan: Fig.3, Col.12 Line 60-66 “performing knowledge distillation comprises: constructing a distillation loss function for training the student network based on the teacher network, wherein the distillation loss function comprises a plurality of loss functions corresponding to one or more layers of the teacher network, and the one or more layers of the teacher network comprise: a weight layer; and an output layer.”)
It would have been obvious to a person of ordinary skill in the art, before the effective filing date of the invention, to modify the teaching of Chen to include the system that inputting the set of tokens to each of a plurality of teacher models and computing at least one of a distillation loss or a contrastive loss using at least one set of logits from one teacher model as taught by Yan. One would have been motivated to make such a combination in order to provide users with efficient and robust Machine learning model that are light-weight model that is not only small in size (meaning less storage footprint and cheaper computational cost) but also can produce accurate results (predictions, classification, etc.) (see Yan Col.4, Line 40-45).
Regarding Claim 7,
Chen, Bui, SUN, Li, ZHAO and Yan teaches all the limitations of Claim 6. Yan further teaches the method wherein:
the distillation loss is computed using the set of logits generated by the predictor (see Yan: Fig.3, Col.12 Line 60-66, “constructing a distillation loss function for training the student network based on the teacher network, wherein the distillation loss function comprises a plurality of loss functions corresponding to one or more layers of the teacher network. In some embodiments, the one or more layers of the teacher network comprise: an embedding layer; a self-attention layer; a feedforward layer; and an output layer.”), and
the set of logits generated by an in-domain teacher model, the in-domain teacher model being the teacher model that is pre- trained in the domain corresponding to the ground-truth domain of the data sample (see Yan: Fig.3, Col.12 Line 60-66 “the plurality of loss functions comprise at least one of the following: a loss function based on a mean-square error of a difference between one or more embedding layers of the student network and the