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 .
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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 01/08/2026 has been entered.
Response to Amendment
The office action is responsive to the amendment filed on 01/08/2026. As directed by the amendments claims 1, 5, 8, 11, 14 and 17 are amended. Claims 1-20 are pending for examination.
Response to Arguments
Regarding the 35 U.S.C § 101 Rejection:
Applicant’s arguments, see Remarks pg. 9-10, filed 01/08/2026, with respect to claims 1-20 have been fully considered and are persuasive. The rejection of claims 1-20 under 35 U.S.C § 101 has been withdrawn.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-3, 8-10 and 14-15 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et al. Meta-Prototypical Learning for Domain-Agnostic Few-Shot Recognition (hereinafter Wang) in view of Finn et al. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (hereinafter Finn) as cited in IDS filed on 10/04/2022.
Regarding claim 1:
Wang teaches A computer-implemented method for domain adaptation comprising; ( Wang pg. 6991, sec: III Meta Prototypical Learning, teaches a method for domain adaptation and pg. 6993, left colm., sec: Experiments, para. 1 teaches “all experiments were implemented with PyTorch” for which a person skilled in the relevant art will recognize PyTorch requires a computer to run).
obtaining a plurality of tasks from a first domain, the plurality of tasks including at least a first task and a second task, (Wang Abstract, teaches a model optimized on data from one domain (first domain) and pg. 6991, sec: III Meta Prototypical Learning, para. 2, teaches obtaining few-shot task sampled from the training set. Where each few shot task (i.e., first task and a second task) can include support set of shots (a few labeled examples) and the query set of queries (data to be recognized), see Fig. 2 and pg. 6690, right colm., para. 1).
training a machine learning system to perform the first task; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-2, teaches performing training with the first task. Specifically Wang teaches “
T
is a few-shot task sampled from the training set ... The inner problem is adapting the parameter of feature encoder
ϕ
for each specific few-shot task”, thus suggesting each training iteration sample one task at a time).
generating a first set of prototypes associated with a first set of classes of the first task; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-2 & Fig. 2, teaches sampling a task ( i.e., first task) and generate class protypes from the support set in the task to classify sample in the query set).
optimizing a first loss output that includes a first task loss, the first task loss being computed based on the first set of prototypes; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query, thus the loss depends on the protypes via classification of query samples).
updating the machine learning system based on the first loss output; (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2, teaches the machine learning system is updated in the “inner update” using gradients from support set S that are derived from the few show task (i.e., first task). Further, Wang pg. 6992, left colm., para. 1-2 teaches the inner updated is based on gradient descent as
PNG
media_image1.png
35
316
media_image1.png
Greyscale
).
training the machine learning system to perform the second task after the machine learning system is updated based on the first loss output; (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2 and Fig. 2, teaches the method “can adapt to different tasks” and teaches performing several inner updates, thus suggesting each training iteration sample one task at a time which includes the “second task” and performs training with the task during the inner update).
generating a second set of prototypes associated with a second set of classes of the second task; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-2 & Fig. 2, teaches sampling a task ( i.e., second task) and generate class protypes from the support set in the task).
optimizing a second loss output that includes a second task loss, the second task loss being computed based on the second set of prototypes; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query, thus the loss depends on the protypes via classification of query samples).
updating the machine learning system based on the second loss output; (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2, teaches the machine learning system is updated in the “inner update” using gradients from support set S that are derived from the few show task (i.e., second task). Further, Wang pg. 6992, left colm., para. 1-2 teaches the inner updated is based on gradient descent as
PNG
media_image1.png
35
316
media_image1.png
Greyscale
).).
performing a meta-update of the machine learning system by adjusting a parameter of the machine learning system to improve generalization of the machine learning system across the plurality of tasks; (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 3; pg. 6692, left colm., para 1; and Fig. 2 teaches performing “meta-update” of the meta-encoder (i.e., machine learning system) by adjusting the parameters of the meta-encoder to “adapt to few-shot tasks from different domains by the traces of the few labeled examples”, thus to improve generalization the meta-encoder across the plurality of tasks (Abstract)).
Wang does not teach or suggest obtaining a new task from a second domain; and fine-tuning the machine learning system with the new task to generate an adapted machine learning system for the second domain.
Nevertheless, Finn teaches the following:
obtaining a new task from a second domain; and ( Finn pg. 3, left colm, para. 2, teaches adapting to a new task and pg. 4, Algorithm 2 MAML for few Shot Supervised Learning, line 8 teaches obtaining a new task
T
i
by Sample datapoint
D
i
'
=
x
j
,
y
j
(i.e. second domain)).
fine-tuning the machine learning system with the new task to generate an adapted machine learning system for the second domain (Fin pg. 2, left colm., sec: 2.1 Meta learning Problem Set Up, para. 1 teaches the goal of meta learning is to train a model that can quickly adapt to a new task and Finn pg. 2, right colm., sec: 2.2. A Model-Agnostic Meta-Learning Algorithm, para. 1 teaches the model is “fine-tuned using a gradient-based learning rule on a new task” drawn from distribution over tasks (i.e., second domain). Algorithm 2 MAML for few Shot Supervised Learning, lines 7-8 teaches how this is achieved, by fine-tuning the model (denoted by
f
) parameters with gradient descent such that the model can “quickly learn a new task from a small amount of new data” ( pg. 1, right col, para. 1, lines 9-12)).
Finn is also in the same field of endeavor as Wang (meta-learning for domain adaptation). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of obtaining new task and fine tuning the machine learning system with the new task, as being disclosed and taught by Finn, in the system taught by Wang to yield the predictable results of providing a “simple model and task-agnostic algorithm for meta-learning that trains a model’s parameters such that a small number of gradient updates will lead to fast learning on a new task” (Finn, pg. 2, left col., para. 1).
Regarding claim 2:
Wang and Finn teach The computer-implemented method of claim 1. Wang specifically teaches wherein: the first task includes a first support set and a first query set; ( Wang Fig. 2 and pg. 6690, right colm., para. 1 teaches each few shot task (i.e., first task) can include support set of shots (a few labeled examples) and the query set of queries (data to be recognized)).
the first set of prototypes is computed using the first support set; and (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-2 & Fig. 2, teaches sampling a task ( i.e., first task) and generate class protypes from the support set in the task to classify sample in the query set).
the first task loss is computed using the first query set ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query).
Regarding claim 3:
Wang and Finn teach The computer-implemented method of claim 1. Wang specifically teaches pg. 6991, sec: III Meta Prototypical Learning, para. 2 and Fig. 2, the method “can adapt to different tasks” and teaches performing several inner updates, thus suggesting each training iteration sample one task at a time which includes the “second task” and performs training with the task during the inner update. Therefore Wang teaches wherein: the second task includes a second support set and a second query set; ( Wang Fig. 2 and pg. 6690, right colm., para. 1 teaches each few shot task (i.e., second task) can include support set of shots (a few labeled examples) and the query set of queries (data to be recognized)).
the second set of prototypes is computed using the second support set; and (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-2 & Fig. 2, teaches sampling a task ( i.e., first task) and generate class protypes from the support set in the task to classify sample in the query set).
the second task loss is computed using the second query set ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query).
Regarding claim 8: is rejected under the same rational of claim 1. Claim 8 only recites the additional elements of One or more non-transitory computer readable storage media storing computer readable data with instructions that when executed by one or more processors cause the one or more processors to perform a method for domain adaptation that comprise... , for which Finn, pg. 11, sec. A.1. Classification teaches “All models were trained for iterations on a single 60000 NVIDIA Pascal Titan X GPU” which inheritably will require a computer readable storage media to operate.
Regarding claim 9: is rejected under the same rational of claim 2.
Regarding claim 10: is rejected under the same rational of claim 3.
Regarding claim 14:
Wang teaches A computer-implemented method for domain adaptation comprising: ( Wang pg. 6991, sec: III Meta Prototypical Learning, teaches a method for domain adaptation and pg. 6993, left colm., sec: Experiments, para. 1 teaches “all experiments were implemented with PyTorch” for which a person skilled in the relevant art will recognize PyTorch requires a computer to run).
obtaining a first task from a plurality of tasks in a source domain, the first task including a first support set and a first query set; (Wang Abstract, teaches a model optimized on data from one domain (source domain) and pg. 6991, sec: III Meta Prototypical Learning, para. 2, teaches obtaining few-shot task sampled from the training set. Where each few shot task (i.e., first task) can include support set of shots (a few labeled examples) and the query set of queries (data to be recognized), see Fig. 2 and pg. 6690, right colm., para. 1).
generating, via a machine learning system, first support output in response to the first support set; ( Applicant specification [0036-37] states “The machine learning system 140 is configured to provide the support output and the query output to the prototype engine [such that] ...generates prototypes for each class of the given task”, thus suggesting the “support output and the query output” being used by the protype engine are embedding produced by the machine learning system for which Wang pg. 6992, sec: B. Meta Update, para.1., teaches “the encoder maps images from support and query set into the same embedding space” hence the embedded support sets can be view as the “support output”).
generating a first set of prototypes for each class of the first task using the first support output; ( Wang pg. 6992, sec: B. Meta Update, para.1, teaches “the encoder maps images from support and query set into the same embedding space. The embeddings of shots are averaged as class prototypes ” thus prototypes for each class is generated from the embedded support sets (i.e., support output)).
computing a first loss output that includes at least a first task loss, the first task loss being computed based on the first set of prototypes and the first query output; (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query, thus the loss depends on the protypes via classification of query samples. In addition, Wang pg. 6992, sec: B. Meta Update, para. 1, function 5 continuing right colm. para 1, and Function 6, teaches the task loss being computed based on the distance of the query embedding (i.e., query output) and the class prototype).
training and updating the machine learning system with respect to remaining tasks of the plurality of tasks; ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2 and Fig. 2, teaches the method “can adapt to different tasks” and teaches performing several inner updates, thus suggesting each training iteration, training will be performed with remaining tasks of the plurality of tasks. In addition, Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2, and Wang pg. 6992, left colm., para. 1-2 teaches the machine learning system is updated in the “inner update” using gradient descent as
PNG
media_image1.png
35
316
media_image1.png
Greyscale
).
performing a meta-update of the machine learning system by adjusting the parameter of the machine learning system to improve generalization of the machine learning system across the plurality of tasks; and (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 3; pg. 6692, left colm., para 1; and Fig. 2 teaches performing “meta-update” of the meta-encoder (i.e., machine learning system) by adjusting the parameters of the meta-encoder to “adapt to few-shot tasks from different domains by the traces of the few labeled examples”, thus to improve generalization the meta-encoder across the plurality of tasks (Abstract)).
Wang does not teach or suggest fine-tuning the machine learning system with a few-shot examples from a new task in a target domain to generate an adapted machine learning system for the target domain.
Nonetheless, Finn teaches the following:
fine-tuning the machine learning system with a few-shot examples from a new task in a target domain to generate an adapted machine learning system for the target domain ( (Fin pg. 2, left colm., sec: 2.1 Meta learning Problem Set Up, para. 1 teaches the goal of meta learning is to train a model that can quickly adapt to a new task only a few datapoints (i.e., few-shot examples) and Finn pg. 2, right colm., sec: 2.2. A Model-Agnostic Meta-Learning Algorithm, para. 1 teaches the model is “fine-tuned using a gradient-based learning rule on a new task” drawn from distribution over tasks (target domain). Algorithm 2 MAML for few Shot Supervised Learning, lines 7-8 teaches how this is achieved, by fine-tuning the model (denoted by
f
) parameters with gradient descent such that the model can “quickly learn a new task from a small amount of new data” ( pg. 1, right col, para. 1, lines 9-12)).
Finn is also in the same field of endeavor as Wang (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of obtaining new task and fine tuning the machine learning system with the new task, as being disclosed and taught by Finn, in the system taught by Wang to yield the predictable results of providing a “simple model and task-agnostic algorithm for meta-learning that trains a model’s parameters such that a small number of gradient updates will lead to fast learning on a new task” (Finn, pg. 2, left col., para. 1).
Regarding claim 15:
Wang, Finn and Snell teach The computer-implemented method of claim 14. Snell specifically teaches wherein the step of training the machine learning system with respect to the remaining tasks of the plurality of tasks further comprises: ( Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2 and Fig. 2, teaches the method “can adapt to different tasks” and teaches performing several inner updates, thus suggesting each training iteration, training will be performed with remaining tasks of the plurality of tasks to generate a “another support output”, “another query output”, “another set of prototypes” for which another loss output will be computed).
generating, via the machine learning system, another support output based on another support set; ( Applicant specification [0036-37] states “The machine learning system 140 is configured to provide the support output and the query output to the prototype engine [such that] ...generates prototypes for each class of the given task”, thus suggesting the “support output and the query output” being used by the protype engine are embedding produced by the machine learning system for which Wang pg. 6992, sec: B. Meta Update, para.1., teaches “the encoder maps images from support and query set into the same embedding space” hence the embedded support sets can be view as the “support output”).
generating another set of prototypes for each class of another task using the another support output; (Snell Algorithm 1, teaches protypes being generated for each class of the first task using the support output).
generating, via the machine learning system, another query output based on another query set; ( Wang pg. 6992, sec: B. Meta Update, para.1, teaches “the encoder maps images from support and query set into the same embedding space. The embeddings of shots are averaged as class prototypes ” thus prototypes for each class is generated from the embedded support sets (i.e., support output)).
computing another loss output, the another loss output including another task loss based on the another set of prototypes and the another query output; and (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query, thus the loss depends on the protypes via classification of query samples. In addition, Wang pg. 6992, sec: B. Meta Update, para. 1, function 5 continuing right colm. para 1, and Function 6, teaches the task loss being computed based on the distance of the query embedding (i.e., query output) and the class prototype).
updating the parameter of the machine learning system based on the another loss output, wherein the another task includes the another support set and the another query set ( A person skilled in the revenant art will recognize that SGD iteratively updates the model (i.e., machine learning system) parameters for which Wang pg. 6991, sec: III Meta Prototypical Learning, para. 2, and Wang pg. 6992, left colm., para. 1-2 teaches the machine learning system is updated in the “inner update” using gradient descent as
PNG
media_image1.png
35
316
media_image1.png
Greyscale
and teaches obtaining few-shot task sampled from the training set. Where each few shot task (i.e., first task) can include support set of shots (a few labeled examples) and the query set of queries (see Fig. 2, pg. 6991, sec: III Meta Prototypical Learning, para. 2, & pg. 6690, right colm., para. 1)).
Claim 4 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Nichol et al. of On First-Order Meta-Learning Algorithms (hereinafter Nichol) as being disclose in the IDS filled on 10/04/2022.
Regarding claim 4:
Wang and Finn teaches The computer-implemented method of claim 1.
Neither Wang or Finn teaches wherein: the machine learning system is trained with the first task a first plurality of times; and the machine learning system is trained with the second task a second plurality of times.
Nonetheless, Nichol teaches the following:
wherein: the machine learning system is trained with the first task a first plurality of times; and
PNG
media_image2.png
130
483
media_image2.png
Greyscale
( Nichol pg. 3, sec. 3 Reptile, para. 1 and Algorithm 1 Reptile (serial version) teaches a neural network model (machine learning system) for which the parameters are initialize and the model is trained multiple iteration (see FOR loop indicated in the above Algorithm) on the sample task (first task)).
the machine learning system is trained with the second task a second plurality of times ( Similar, the Reptile algorithm from Nichol pg. 3, sec. 3 Reptile, para. 1 and Algorithm 1 Reptile (serial version) can be used to trained the neural network model (machine learning system) with the second task (sample task) during multiple iterations).
Nichol is also in the same field of endeavor as Wang and Finn (machine learning – meta learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of updating the machine learning system based on its loss output, as being disclosed and taught by Nichol, in the system taught by Wang and Finn to yield the predictable results of providing “a theoretical analysis that applies to both first -order MAML and Reptile, showing that they both optimize for within-task generalization” ( Nichol, pg. 2, 3rd bullet point).
Claims 5, 11 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Bergman et al. Classification – Based Anomaly Detection for General Data (hereinafter Bergman) as being disclose in the IDS filled on 10/04/2022.
Regarding claim 5:
Wang and Finn teaches The computer-implemented method of claim 1. Finn specifically teaches obtaining sample...data associated with the new task ( Finn pg. 4, Algorithm 2 MAML for few Shot Supervised Learning, line 8 teaches obtaining a new task
T
i
by Sample datapoint
D
i
'
=
x
j
,
y
j
).
However, neither Wang or Finn disclose obtaining sample input data associated with the new task; generating, via the machine learning system, sample output data in response to the sample input data; generating an anomaly score for each of the sample input data based on the sample output data; and indicating whether a particular sample is anomalous when an associated anomaly score differs from an expected value beyond a threshold.
Nevertheless Bergman teaches the following:
further comprising: obtaining sample input data associated with the new task; (Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches Test Sample
x
( sample input data)).
generating, via the machine learning system, sample output data in response to the sample input data; (Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches the model generates a likelihood of predicting the correct transformation which can be seen as the sample output data).
generating an anomaly score for each of the sample input data based on the sample output data; and ( Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches input Test Sample
x
( sample input data) for which an anomaly score for each sample of input data and the likelihood ( output data) is being generated denoted as
S
c
o
r
e
(
x
)
).
indicating whether a particular sample is anomalous when an associated anomaly score differs from an expected value beyond a threshold ( Bergman teaches pg. 7, sec. 5.2 Tabular Data Experiment, Implementation of GOAD, teaches “the decision threshold value is chosen to result in the correct number of anomalies e.g. if the test set contains
N
a
anomalies, the threshold is selected so that the highest
N
a
scoring examples are classified as anomalies”).
Bergman is also in the same field of endeavor as Wang and Finn (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of obtaining input data that is used to generate output data and indicating wheatear a particular sample is anomalous when it differs a threshold value as being disclosed and taught by Bergman, in the system taught by Wang and Finn to yield the predictable results of improving over the state of the art, by providing a method for detecting anomalies for general data, by training a classifier on a set of random auxiliary task, the method presented does not requires knowledge of the data domain, doing it’s able to generate an arbitrary number of random tasks ( Bergman pg. 9, sec. Conclusion).
Regarding claim 11:
Wang and Finn teach The one or more non-transitory computer readable storage media of claim 8.
Finn specifically teaches further comprising: obtaining sample...data associated with the new task ( Finn pg. 4, Algorithm 2 MAML for few Shot Supervised Learning, line 8 teaches obtaining a new task
T
i
by Sample datapoint
D
i
'
=
x
j
,
y
j
).
However, neither Wang or Finn disclose obtaining sample input data associated with the new task; generating, via the machine learning system, sample output data in response to the sample input data; generating an anomaly score for each of the sample output data; and indicating whether or not each of the sample input data is anomalous or non-anomalous by comparing each anomaly score with a threshold value.
Nevertheless Bergman teaches the following:
obtaining sample input data associated with the new task; (Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches Test Sample
x
( sample input data)).
generating, via the adapted machine learning system, sample output data in response to the sample input data; (Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches the model generates a likelihood of predicting the correct transformation which can be seen as the sample output data).
generating an anomaly score for each of the sample output data; and ( Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches generating an anomaly score for each sample of input data and the likelihood ( output data) denoted as
S
c
o
r
e
(
x
)
).
indicating whether or not each of the sample input data is anomalous or non-anomalous by comparing each anomaly score with a threshold value ( Bergman teaches pg. 7, sec. 5.2 Tabular Data Experiment, Implementation of GOAD, teaches “the decision threshold value is chosen to result in the correct number of anomalies e.g. if the test set contains
N
a
anomalies, the threshold is selected so that the highest
N
a
scoring examples are classified as anomalies”).
Regarding claim 17:
Wang and Finn teach The computer-implemented method of claim 14. Finn specifically teaches further comprising: obtaining the new task, the new task including (i) the few-shot examples and (ii) samples; (Finn algorithm 2, lines 3-5 and 8 teaches obtaining new task
T
i
(line 3-4), the new task including Sample
K
, line 5, (i.e., few shot example) and Sample Datapoint
D
i
'
, line 8, (i.e., samples)).
Neither Wang and Finn teach generating, via the adapted machine learning system, sample output data in response to the samples; generating an anomaly score for each of the samples; and indicating whether a particular sample is anomalous when an associated anomaly score differs from an expected value beyond a threshold.
Nevertheless, Bergman teaches the following:
generating, via the machine learning system, sample output data in response to the samples; (Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches the model generates a likelihood of predicting the correct transformation which can be seen as the sample output data).
generating an anomaly score for each of the samples; and ( Bergman pg. 4, Algorithm 2: GOAD: Evaluation Algorithm, teaches input Test Sample
x
( sample input data) for which an anomaly score for each sample is being generated denoted as
S
c
o
r
e
(
x
)
).
indicating whether a particular sample is anomalous when an associated anomaly score differs from an expected value beyond a threshold ( Bergman teaches pg. 7, sec. 5.2 Tabular Data Experiment, Implementation of GOAD, teaches “the decision threshold value is chosen to result in the correct number of anomalies e.g. if the test set contains
N
a
anomalies, the threshold is selected so that the highest
N
a
scoring examples are classified as anomalies”).
Bergman is also in the same field of endeavor as Wang and Finn (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of generating output data, anomaly score and indicating if a sample is anomalous when associated with an anomaly score as being disclosed and taught by Bergman, in the system taught by Wang and Finn to yield the predictable results of improving over the state of the art, by providing a method for detecting anomalies for general data, by training a classifier on a set of random auxiliary task, the method presented does not requires knowledge of the data domain, doing it’s able to generate an arbitrary number of random tasks ( Bergman pg. 9, sec. Conclusion).
Claims 6, 12, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Snell et al. Prototypical Networks for Few-shot Learning as cited in IDS filed on 10/04/2022.
Regarding claim 6:
Wang and Finn teach The computer-implemented method of claim 1.
Wang specifically teaches generating, via the machine learning system, few-shot output data in response to the few-shot examples ( Wang pg. 6993, left colm., sec: A. Datasets, Protocols, and Implementation Details, para. 1, teaches “sample few-shot tasks” (i.e., few shot examples) from different domains for evaluation”. Further, as stated in applicant specification paragraph [0046] "the domain adapting framework 130 includes a few-shot learning method that classifies samples based on distance on the embedding space". Thus, this suggest the few shot output can be seen as the predicted class label for each query point for which Wang pg. 6992,sec: B Meta Update, para. 1, and Function 5 teaches the few-shot output is being generated based on a distribution over classes for a query point x based on a SoftMax over distances);
Finn further teaches the following:
wherein the step of fine-tuning the machine learning system with the new task comprises: ( Finn pg. 4, Algorithm 2 MAML for few Shot Supervised Learning, lines 7-8 teaches fine-tuning a model (denoted by
f
) parameters with gradient descent (line 7) such that the model can “quickly learn a new task from a small amount of new data” ( pg. 1, right col, para. 1, lines 9-12)).
obtaining a few-shot examples associated with the new task; (Finn, pg. 2, right col., para. 1, teaches model obtains
K
samples (few-shot examples) and “feedback from the corresponding loss” from the new task
T
i
).
Neither Wang or Finn teaches generating, via the machine learning system, few-shot output data in response to the few- shot examples; generating a new set of prototypes associated with a new set of classes of the new task; optimizing a new loss output that includes a new task loss, the new task loss being based on the new set of prototypes and the few-shot output data; and updating the machine learning system based on the new loss output.
Nonetheless, Snell teaches the following:
obtaining a few-shot examples ; ( Snell pg. 2, sec. 2.1 Notation, teaches “In few-shot classification we are given a small support set of
N
labeled examples
S
”, which represent the few-shot examples).
generating, via the machine learning system, few-shot output data in response to the few-shot examples; ( Applicant specification paragraph [0046] states " the domain adapting framework 130 includes a few-shot learning method that classifies samples based on distance on the embedding space". Thus, this suggest the few shot output can be seen as the predicted class label for each query point for which Snell pg. 2, sec. 2.2 Model, para. 3, teaches the few-shot output is being generated based on “a distribution over classes for a query point x based on a SoftMax over distances”).
generating a new set of prototypes associated with a new set of classes of the new task; ( Snell pg. 2, sec: 2.2 Model, para. 3, teaches “Training episodes [tasks] are formed by randomly selecting a subset of classes from the training set, then choosing a subset of examples within each class to act as the support set and a subset of the remainder to serve as query points” and Algorithm 1 teaches RANDOMSAMPLE (S,N) denotes a set of N elements chosen uniformly at random from set S, without replacement. Thus, under broadest reasonable interpretation (BRI) each execution/run of algorithm 1 correspond to a different episode (i.e., new task) for which new prototype will be generated by function
PNG
media_image3.png
56
220
media_image3.png
Greyscale
. That is, each execution/run of Algorithm 1, will produce a new task, a new support set , a new query set, a new prototype and a new loss).
optimizing a new loss output that includes a new task loss, the new task loss being based on the new set of prototypes and the few-shot output data; and
PNG
media_image4.png
547
866
media_image4.png
Greyscale
( Snell Algorithm 1 teaches a new loss is initialize (line 10) for a training episode (i.e., task) and pg. 2, sec: 2.2. model, para. 3 teaches the loss being is optimized by implementing stochastic gradient descent (SDG) to minimize the negative log probability. Further, Algorithm 1, line 13 teaches the new task loss J being based on the new set of prototype
c
k
and the few-shot output data)
updating the machine learning system based on the new loss output ( Snell teaches pg. 2, sec: 2.2. model, para. 3 teaches the loss being is optimized by implementing (SDG). A person skilled in the revenant art will recognize that SGD iteratively updates the model (i.e., machine learning system) parameters, hence update the model, to minimize the loss function).
Snell is also in the same field of endeavor as Wang and Finn (machine learning specifically meta-learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of few shot examples and samples being obtained and generated as well as new protypes being generated from new task as shown in Algorithm 1, as being disclosed and taught by Snell, in the system taught by Wang and Finn to yield the predictable results of providing a “prototypical network” for the problem of few-shot classification, that “can yield substantial improvements over recent approaches involving complicated architectural choices and meta-learning” ( Snell Abstract).
Regarding claim 12: is rejected under the same rational of claim 6.
Regarding claim 19:
Wang and Finn teaches The computer-implemented method of claim 14.Neither Wang or Finn explicitly teaches wherein each prototype of the first set of prototypes is a class centroid.
Nevertheless, Snell teaches the following: Snell specifically teaches wherein each prototype of the first set of prototypes is a class centroid ( Snell, pg. 2 equation 1, and Fig. 1 teaches each prototype includes a class centroid
c
k
, ).
Claims 7 and 13 is rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Hendrycks et al. Depp Anomaly Detection With Outlier Exposure (hereinafter Hendrycks) as being disclose in the IDS filled on 10/04/2022.
Regarding claim 7:
Wang and Fin teach The computer-implemented method of claim 1.
Wang and Finn do not suggest further comprising: computing an outlier loss based on outlier data that does not belong to a normal distribution of the first task, wherein the first loss output is computed based on the first task loss and the outlier loss.
Nevertheless, Hendrycks teaches the following:
further comprising: computing an outlier loss based on outlier data that does not belong to a normal distribution of the first task, wherein the first loss output is computed based on the first task loss and the outlier loss ( Hendrycks pg. 1, Introduction, teaches “training anomaly detectors against an auxiliary dataset of outliers, an approach we call Outlier Exposure (OE). This enables anomaly detectors to generalize and detect unseen anomalies” and pg. 3, para. 1, teaches “Outlier Exposure can be applied with many types of data and original tasks” and the specific formulation of the outlier loss
L
O
E
“is a design choice, and depends on the task at hand and the (OOD) detector used”. Further, Hendrycks pg. 2, sec. 4.3 Multiclass Classification, teaches the outlier loss
L
O
E
is defined as a cross entropy to the uniform distribution and its computed using the following equation:
PNG
media_image5.png
25
386
media_image5.png
Greyscale
Where “
H
is the cross entropy and
U
is the uniform distribution over
K
class” and
D
o
u
t
O
E
is Outlier Exposure (OE) dataset ( i.e., the outlier data, see Hendrycks, pg. 2, sec. 3 Outlier Exposure, para. 1)).
Hendrycks is also in the same field of endeavor as Wang and Finn (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of an outlier exposure loss
L
O
E
of Hendrycks in combination with the prototype loss of Snell in order to compute a loss output. The motivation to do so, is that Hendrycks provides the means to improve deep anomaly detection by providing an approach called Outlier Exposure (OE) that enables anomaly detector to generalize and detect unseen anomaly ( Hendrycks Abstract).
Regarding claim 13: is rejected under the same rational of claim 7.
Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn, Snell in further view of Venkataramani et al. US 20200118043 A1 (hereinafter Venkataramani).
Regarding claim 16:
Wang, Finn and Snell teach The computer-implemented method of claim 14. Wang specifically teaches generating, via the machine learning system, few-shot output in response to the few shot examples; ( Wang pg. 6993, left colm., sec: A. Datasets, Protocols, and Implementation Details, para. 1, teaches “sample few-shot tasks” (i.e., few shot examples) from different domains for evaluation”. Further, as stated in applicant specification paragraph [0046] "the domain adapting framework 130 includes a few-shot learning method that classifies samples based on distance on the embedding space". Thus, this suggest the few shot output can be seen as the predicted class label for each query point for which Wang pg. 6992,sec: B Meta Update, para. 1, and Function 5 teaches the few-shot output is being generated based on a distribution over classes for a query point x based on a SoftMax over distances).
While Wang does not specifically teaches wherein the step of fine-tuning the machine learning system with the few-shot examples from the new task in the target domain further comprises: generating new prototypes for each class of the new task using the few-shot output; computing a new task loss based on the new prototypes and the few-shot output; updating the parameter of the machine learning system based on the new task loss; and deploying the machine learning system in the target domain.
Finn overcome this deficiencies and teaches the following:
Finn teaches a model obtains
K
samples (few-shot examples) and “feedback from the corresponding loss” from the new task
T
i
(Finn, pg. 2, right col., para. 1) and further teaches wherein the step of fine-tuning the machine learning system with the few-shot examples from the new task in the target domain further comprises: (Fin pg. 2, left colm., sec: 2.1 Meta learning Problem Set Up, para. 1 teaches the goal of meta learning is to train a model that can quickly adapt to a new task for only a few datapoints (i.e., few-shot examples) and Finn pg. 2, right colm., sec: 2.2. A Model-Agnostic Meta-Learning Algorithm, para. 1 teaches the model is “fine-tuned using a gradient-based learning rule on a new task” drawn from distribution over tasks (target domain). Algorithm 2 MAML for few Shot Supervised Learning, lines 7-8 teaches how this is achieved, by fine-tuning the model (denoted by
f
) parameters with gradient descent such that the model can “quickly learn a new task from a small amount of new data” ( pg. 1, right col, para. 1, lines 9-12)).
updating the parameter of the machine learning system based on the new task loss; (Finn pg. 3, left colm, para 2 and pg. 4, Algorithm 2 MAML for few Shot Supervised Learning, line 10 teaches updating the parameters
θ
of the machine learning system based on the new task loss
L
T
i
using gradient descent updates).
Neither Wang or Finn specifically teaches or suggest generating, via the machine learning system, few-shot output in response to the few shot examples; generating new prototypes for each class of the new task using the few-shot output; computing a new task loss based on the new prototypes and the few-shot output; and deploying the machine learning system in the target domain.
Nonetheless Snell teaches the following:
generating, via the machine learning system, few-shot output in response to the few shot examples; ( Applicant specification paragraph [0046] states " the domain adapting framework 130 includes a few-shot learning method that classifies samples based on distance on the embedding space". Thus, this suggest the few shot output can be seen as the predicted class label for each query point for which Snell pg. 2, sec. 2.2 Model, para. 3, teaches the few-shot output is being generated based on “a distribution over classes for a query point x based on a SoftMax over distances”).
generating new prototypes for each class of the new task using the few-shot output; ( Snell pg. 2, sec: 2.2 Model, para. 3, teaches “Training episodes [tasks] are formed by randomly selecting a subset of classes from the training set, then choosing a subset of examples within each class to act as the support set and a subset of the remainder to serve as query points” and Algorithm 1 teaches RANDOMSAMPLE (S,N) denotes a set of N elements chosen uniformly at random from set S, without replacement. Thus, under broadest reasonable interpretation (BRI) each execution/run of algorithm 1 correspond to a different episode (i.e., new task) for which new prototype will be generated by function
PNG
media_image3.png
56
220
media_image3.png
Greyscale
. That is each execution/run of Algorithm 1, will produce a new task, a new support set , a new query set, a new prototype and a new loss).
computing a new task loss based on the new prototypes and the few-shot output; ( Snell pg. 2, sec. 2.2 Model, para. 3, and Algorithm 1, teaches computing a loss output
J
ϕ
from the protypes
c
k
and the query embedding
f
ϕ
x
(i.e., query output)).
Snell is also in the same field of endeavor as Wang and Finn (machine learning specifically meta-learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of few shot examples and samples being obtained and generated as well as new protypes being generated from new task as shown in Algorithm 1, as being disclosed and taught by Snell, in the system taught by Wang and Finn to yield the predictable results of providing a “prototypical network” for the problem of few-shot classification, that “can yield substantial improvements over recent approaches involving complicated architectural choices and meta-learning” ( Snell Abstract).
Neither Wang, Finn or Snell disclose deploying the machine learning system in the target domain.
However, Venkataramani teaches:
and deploying the machine learning system in the target domain ( Venkataramani [0077] teaches “the machine-learnt model 106 is deployed in new target domains”).
Venkataramani is also in the same field of endeavor as Wang, Finn and Snell (machine learning – few shot learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of updating and fine-tuning the machine learning system, as being disclosed and taught by Venkataramani, in the system taught by Wang, Finn and Snell to yield the predictable results of facilitating “enhanced memory augmented continuous learning for adapting a machine-learnt model to a new domain to deliver better performance with a relatively small set of samples” (Venkataramani [0014]).
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Hendrycks.
Regarding claim 18:
Wang and Finn teach The computer-implemented method of claim 14.
Wang and Finn do not suggest further comprising: computing an outlier loss based on outlier data that does not belong to a normal distribution of the first task, wherein the first loss output includes the first task loss and the outlier loss.
Nevertheless, Hendrycks teaches the following:
further comprising: computing an outlier loss based on outlier data that does not belong to a normal distribution of the first task, wherein the first loss output includes the first task loss and the outlier loss ( Hendrycks pg. 1, Introduction, teaches “training anomaly detectors against an auxiliary dataset of outliers, an approach we call Outlier Exposure (OE). This enables anomaly detectors to generalize and detect unseen anomalies” and pg. 3, para. 1, teaches “Outlier Exposure can be applied with many types of data and original tasks” and the specific formulation of the outlier loss
L
O
E
“is a design choice, and depends on the task at hand and the (OOD) detector used”. Further, Hendrycks pg. 2, sec. 4.3 Multiclass Classification, teaches the outlier loss
L
O
E
is defined as a cross entropy to the uniform distribution and its computed using the following equation:
PNG
media_image5.png
25
386
media_image5.png
Greyscale
Where “
H
is the cross entropy and
U
is the uniform distribution over
K
class” and
D
o
u
t
O
E
is Outlier Exposure (OE) dataset ( such as the outlier data, see Hendrycks, pg. 2, sec. 3 Outlier Exposure, para. 1)).
Hendrycks is also in the same field of endeavor as Wang and Finn (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of computing outlier loss, as being disclosed and taught by Hendrycks, in the system taught Wang and Finn to yield the predictable results of improve deep anomaly detection by providing an approach called Outlier Exposure (OE) that enables anomaly detector to generalize and detect unseen anomaly ( Hendrycks Abstract).
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Wang, Finn in further view of Li et al. US 20210334664 A1 (hereinafter Li).
Regarding claim 20:
Wang and Finn teach The computer-implemented method of claim 14. Snell specifically teaches wherein: the first task includes first metadata; and ( Wang pg. 6991, sec: III Meta Prototypical Learning, right colm., para 1, teaches few-shot classification, which a person ordinary in the relevant art will recognize that it inheritably involves the use of tasks and teaches “ labeled samples in
S
are used to generate class porotype to classify sample in
Q
” this suggest the support and query set which are part of the “first task” (according to para. [0006] of the specification) include metadata in the form of labels).
the first task loss is computed (Wang pg. 6991, sec: III Meta Prototypical Learning, para. 1-3 & Fig. 2, teaches several steps of inner updates (i.e., optimization) is used to construct (i.e., computing) the loss based on the classification results of samples in the query).
Neither Wang and Finn suggest the loss is computed using the metadata as ground truth data. Nevertheless, Li teaches the following:
the task loss is computed using the first metadata as first ground truth data ( Li [0029] teaches “the ground truth data is representative of metadata” and [0049] teaches “to determine the loss function , the loss module compares the feature probability distribution generated for an instance of new domain data to corresponding ground truth data” which suggest the loss is computed/determined using the metadata as ground truth data).
Li is also in the same field of endeavor as Wang and Finn (machine learning). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to include the functionality of computing the loss based on the metadata as ground truth data as being disclosed and taught by Li, in the system taught by Wang and Finn to yield the predictable results of “reducing an amount of computational and network resources used in training a model” by reducing the training data “such that the techniques described herein do not require vast amounts (e.g., millions) of annotated training data samples to generate a reliable model” ( Li [0024])).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GISEL G FACCENDA whose telephone number is (703)756-1919. The examiner can normally be reached Monday - Friday 8:00 am - 4:00 pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abdullah Al Kawsar can be reached at (571) 270-3169. 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.
/G.G.F./Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights