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 .
Response to Arguments
Applicant’s argument filed 03/16/2026 have been fully considered but they are not persuasive.
Applicant’s Argument: On page 8-12 of Applicant’s response to rejections under 35 U.S.C. 101, applicant states that claim 1 as a whole is not directed to a mental process and/or a mathematical concept. Applicant argues “claim 1 does not recite mathematical calculation as an abstract idea. Instead, claim 1 recites retraining the classifier using prototype vectors as target, which is a controller level learning operation performed by a machine learning system and cannot be reduced to mere calculation.” The Specification discloses that output vectors are generated in a hyperdimensional embedding space. Claim 1 is also analogous to Ex Parte Desjardins because it is directed to an improvement in method for continual training of a classifier to address the issue of catastrophic forgetting.
Examiner’s Response: Applicant’s argument is not persuasive. During examination, the examiner should analyze the "improvements" consideration by evaluating the specification and the claims to ensure that a technical explanation of the asserted improvement is present in the specification, and that the claim reflects the asserted improvement (see MPEP §2106.05(a)). The MPEP (§2106.05(a)(II)) also warns, “it is important to keep in mind that an improvement in the abstract idea itself (e.g. a recited fundamental economic concept) is not an improvement in technology.” Here, the alleged improvement in the form of “retraining the classifier using the one or more second training datasets and the first training dataset by minimizing a distance between the one or more second output vectors and the one or more prototype vectors” is not a technological improvement in classifier training. Training or re-training a machine learning model is generic computer process and the step of minimizing the distance between a training dataset and a new dataset is mere instructions for using a computer as a tool to perform a mathematical calculation. The use of a computer or other machinery in its ordinary capacity does not integrate a judicial exception into a practical application or provide significantly more. Additional, claim 1 does not recite a hyperdimensional embedding space nor retraining over a number of iterations. Thus, these limitations are not considered under the subject matter eligibility analysis.
Applicant’s Argument: On page 12-14 of Applicant’s response to rejections under 35 U.S.C. 101, applicant states that claim 1 integrates the abstract idea into a practical application because the claim as a whole recites how model parameters and memory structures are updated with respect to novel classes.
Examiner’s Response: Applicant’s argument is not persuasive. An important consideration in determining whether a claim improves technology is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome, as opposed to merely claiming the idea of a solution or outcome (see MPEP 2106.05(a)). The amended claims do not provide sufficient details to describe any technological improvement. If the specifications explicitly set forth an improvement but in a conclusory manner (see MPEP 2106.04(d)(1): a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology.
Applicant’s Argument: On page 14-17 of Applicant’s response to rejections under 35 U.S.C. 103, applicant states that Nishimaki does not teach retraining a classifier during incremental sessions. Applicant also argues that Cheraghian fails to teach retraining the classifier by minimizing a distance between the output vectors and the prototype vectors. The losses disclosed by Cheraghian is not an optimization objective for retraining a classifier. Additionally, Cheraghian discloses semantic word vector, which is different from prototype vector.
Examiner’s Response: Applicant’s argument is not persuasive. The claim limitation “retraining the classifier using the one or more second training datasets and the first training dataset by minimizing a distance between the one or more second output vectors and the one or more prototype vectors” is recited in a broad manner and there are no details as to what constitutes as “retraining the classifier” and “minimizing a distance between the output vector and the prototype vector”. In the previous Office Action (dated 12/15/2025; pg. 19), Examiner states that the Nishimaki reference discloses the retraining of the classifier by meta-learning. Nishimaki (par. 49) discloses the GAT learns the dependency between the basic dataset and the new dataset by meta-learning. Meta-learning by reconstructing weights is a form of training to generate optimized weights for new tasks. In Figure 2 of Nishimaki, stage 3 consists of multiple sessions of inputting new dataset into the novel learning module for meta-learning and each session can be considered a training iteration, where the model learns to classify new data with very limited data samples. The parameters of the GAT (Nishimaki, par. 45) is optimized and updated for each episode. Thus, Nishimaki does teach retraining of a classifier.
It is not clear why semantic word vector cannot be used to teach prototype vector. From claim 1, “prototype vectors” are associated with base classes and there are no additional details to further limit the scope of the definition of the claim element. Cheraghian (pg. 3-4, Section 3.2, par.1) discloses that the feature representations of the first task are associated with semantic representations. The first task consists of many training samples that are associated with a base class (Cheraghian, pg. 3, Section 3.1, par. 1).
Cheraghian (pg. 3, Section 3.1, par. 1) discloses the model is incrementally trained on the training samples for the t-th session. Thus, Cheraghian teaches that a model is trained and updated for a number of training iterations. Cheraghian (pg. 6, “Implementation details”) discloses the experiment uses Adam optimizer, which is an optimization algorithm to train models by updating the model parameters iteratively based on training data. Thus, an optimization objective is used to minimize the loss function. Cheraghian (pg. 3-4, Section 3.2, par.1) discloses computing the cosine distance between project features and the semantic representations for both the base and novel tasks. The distillation loss is computed based on the output of the classifier before adding the novel task and the output of the classifier after adding the novel task. Therefore, the optimization of the distillation loss function is used to minimize the cosine similarity between representations from the base and novel classes.
Claim Rejections - 35 USC § 101
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 1-18 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding Claim 1:
Subject Matter Eligibility Analysis Step 1:
Claim 1 recites “A method for continual training of a classifier, ... , the method comprising” and is thus a process, one of the four statutory categories of patentable subject matter.
Subject Matter Eligibility Analysis Step 2A Prong 1:
“using a set of first output vectors ” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. judgement)
“” (a mathematical calculation; see par. 36 of the Specification)
“determining a set of updated prototype vectors indicative of the first training dataset and the one or more second training datasets” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. judgement)
Claim 1 therefore recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
“... the classifier comprising a controller and an explicit memory ...” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“pre-training the classifier using a first training dataset comprising first data samples of a set of one or more associated base classes” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
"” (This step is directed to data gathering, which is understood to be insignificant extra solution activity - see MPEP 2106.05(g))
“storing the one or more prototype vectors in the explicit memory” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity - see MPEP 2106.05(g))
"receiving one or more second training datasets, each comprising second data samples of a set of one or more associated novel classes” (This step is directed to data gathering, which is understood to be insignificant extra solution activity - see MPEP 2106.05(g))
“adding to the explicit memory one or more second output vectors indicative of the set of one or more associated novel classes, in response to providing the one or more second training datasets to the classifier” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity - see MPEP 2106.05(g))
“retraining the classifier using the one or more second training datasets and the first training dataset by ” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“updating the explicit memory with the set of updated prototype vectors” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity - see MPEP 2106.05(g))
The additional elements as disclosed above alone or in combination do not integrate the judicial exception into practical application as they are mere insignificant extra solution activity in combination of generic computer functions being implemented with generic computer elements in a high level of generality to perform the disclosed abstract idea above. Therefore, Claim 1 is directed to the abstract idea.
Subject Matter Eligibility Analysis Step 2B:
“... the classifier comprising a controller and an explicit memory ...” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“pre-training the classifier using a first training dataset comprising first data samples of a set of one or more associated base classes” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
"” (This step is directed to transmitting or receiving information, which is understood to be insignificant extra solution activity and well understood, routine and conventional activity of transmitting and receiving data as identified by the court - see MPEP 2106.05(d))
“storing the one or more prototype vectors in the explicit memory” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity and well understood, routine and conventional activity of storing and retrieving information in memory as identified by the court - see MPEP 2106.05(d))
"receiving one or more second training datasets, each comprising second data samples of a set of one or more associated novel classes” (This step is directed to transmitting or receiving information, which is understood to be insignificant extra solution activity and well understood, routine and conventional activity of transmitting and receiving data as identified by the court - see MPEP 2106.05(d))
“adding to the explicit memory one or more second output vectors indicative of the set of one or more associated novel classes, in response to providing the one or more second training datasets to the classifier” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity and well understood, routine and conventional activity of storing and retrieving information in memory as identified by the court - see MPEP 2106.05(d))
“retraining the classifier using the one or more second training datasets and the first training dataset by ” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“updating the explicit memory with the set of updated prototype vectors” (This step is directed to storing data in memory, which is understood to be insignificant extra solution activity and well understood, routine and conventional activity of storing and retrieving information in memory as identified by the court - see MPEP 2106.05(d))
The additional elements as disclosed above alone or in combination do not recite significantly more than the abstract idea itself as they are mere insignificant extra solution activity in combination of generic computer functions being implemented with generic computer elements in a high level of generality to perform the disclosed abstract idea above. Therefore, Claim 1 is subject-matter ineligible.
Regarding Claim 2:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the controller comprises a feature extractor and a classification head” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 3:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the feature extractor receives the first data sample and provides a first extracted feature vector, and wherein the classification head receives the first extracted feature vector and provides the set of first output vectors to indicate the one or more associated base classes of the first data sample” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of transmitting and receiving data as identified by the court (2106.05(d) in step 2B))
Regarding Claim 4:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the controller is a convolutional neural network controller comprising multiple nonlinear layers for the feature extractor and the classification head being an output layer of the convolutional neural network” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 5:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“inferring the feature extractor using the one or more second training datasets and ” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. judgement)
“” (a mathematical calculation; see par. 36 of the Specification)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“one or more second training datasets” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of storing and retrieving information in memory as identified by the court (2106.05(d) in step 2B))
“training the classification head using all stored extracted feature vectors” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“wherein the training of the classification head is performed ” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 6:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. judgement)
“set of updated prototype vectors is minimized by using:
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
, where
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
, where ki* is the set of updated prototype vectors in association with an ith class, and ai is an associated extracted feature vector of the classification head for the ith class” (a mathematical calculation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the multiple nonlinear layers with parameter set θ1 and the classification head with parameter set θ2 produce the one or more second output vectors and
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
f” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 7:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the one or more prototype vectors p* •* is defined for the ith class as follows: p*i=
g
θ
2
T
(
a
i
)
” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 8:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated quasi-orthogonal prototype vectors and resultantly ” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. judgement)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“understood, routine and conventional activity of storing and retrieving information in memory as identified by the court (2106.05(d) in step 2B))
Regarding Claim 9:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“wherein determining the quasi-orthogonal prototype vectors comprises backpropagation using a loss function as follows:
PNG
media_image1.png
194
620
media_image1.png
Greyscale
--
where
k
i
(
u
)
is an quasi-orthogonal prototype vector obtained in iteration number u for the ith class, and pi is a prototype vector stored in the explicit memory in association with the ith class, and sh denotes a soft hamming distance, where the initial quasi-orthogonalized prototypes are given by
k
i
(
0
)
=
p
i
” (a mathematical calculation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B: None
Regarding Claim 10:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“providing an activation memory for accumulating the extracted feature vectors of the feature extractor” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of storing and retrieving information in memory as identified by the court (2106.05(d) in step 2B))
“wherein the classification head is configured to receive an input extracted feature vector from the activation memory” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of transmitting and receiving data as identified by the court (2106.05(d) in step 2B))
Regarding Claim 11:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“performing a similarity search between the query vector and the one or more prototype vectors in the explicit memory for determining a class represented by the query vector” (a mathematical calculation; Determining Euclidean distance or cosine similarity)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“receiving a query vector” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of transmitting and receiving data as identified by the court (2106.05(d) in step 2B))
Regarding Claim 12:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“providing an in-memory computing core comprising a crossbar array structure comprising row lines and column lines and resistive memory elements coupled between the row lines and the column lines at junctions formed by the row and column lines” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“programming the resistive memory elements of each column line to represent values of the one or more prototype vectors” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“inputting to the crossbar array the query vector for performing the similarity search” (This step is directed to data gathering, which is understood to be insignificant extra solution activity (2106.05(g) in step 2A prong 2) and well understood, routine and conventional activity of transmitting and receiving data as identified by the court (2106.05(d) in step 2B))
Regarding Claim 13:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the retraining the classifier is a few shot learning performed upon the set of one or more associated novel classes” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 14:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the retraining the classifier comprises retraining a part of the classifier and freezing another pretrained part of the classifier” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 15:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the prototype vector comprises a set of elements each indicating a probability that a respective class of the one or more associated base classes is a class of the one or more prototype vectors” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 16:
The claim recites an article of manufacture that performs the method as described in claim 1. Therefore, claim 16 is rejected for the same reasons as disclosed for claim 1. The limitations for additional elements of claim 16 are analyzed below.
Subject Matter Eligibility Analysis Step 2A Prong 1:
Please see Step 2A Prong 1 analysis of claim 1
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“A computer program product comprising a computer-readable storage medium having computer-readable program code embodied therewith, wherein the computer-readable program code when called by a processor causes the processor to” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 17:
The claim recites a machine that performs the method as described in claim 1. Therefore, claim 17 is rejected for the same reasons as disclosed for claim 1. The limitations for additional elements of claim 17 are analyzed below.
Subject Matter Eligibility Analysis Step 2A Prong 1:
Please see Step 2A Prong 1 analysis of claim 1
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“A computer system for continual training of a classifier, the classifier comprising a controller and an explicit memory, the computer system comprising a processor and a computer-readable storage medium having computer-readable program code embodied therewith, wherein the computer-readable program code when called by the processor causes the processor to” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 18:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“... perform a vector matrix multiplication at the crossbar array for computing a similarity of the query vector with the one or more prototype vectors, thereby determining a class of the query vector” (a mathematical calculation; determining cosine similarity)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“a crossbar array structure comprising row lines and column lines and resistive memory elements coupled between the row lines and the column lines at junctions formed by the row and column lines, the resistive memory elements of each column line representing values of a respective prototype vector, the crossbar array being configured to receive elements of a query vector through the row lines respectively ...” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-4, 11, and 13-17 are rejected under 35 U.S.C. 103 as being unpatentable over Nishimaki (US20250200397A1) in view of Cheraghian, “Semantic-aware Knowledge Distillation for Few-Shot Class-Incremental Learning”.
Regarding claim 1, Nishimaki teaches:
“A method for continual training of a classifier, the classifier comprising a controller and an explicit memory, the method comprising” (abstract, [0035, 0088], A machine learning apparatus for classifying a novel and base dataset is trained with continual learning. The apparatus consists of a feature extraction unit (controller) and a graph attention network that learns in an episodic format, which implies a memory that stores intermediate outputs. The apparatus can be implemented by a memory.)
“pre-training the classifier using a first training dataset comprising first data samples of a set of one or more associated base classes” ([0034], The model is pre-trained with a large amount of base class datasets that includes a plurality of different classes.)
“using a set of first output vectors provided by the controller, in response to the controller receiving the first training dataset, to determine one or more prototype vectors indicative of the set of one or more associated base classes” ([0034-0038], The feature extractor is pre-trained with the basic dataset to output the base class classification weight vector (first output vectors). In stage 2, the base class classification weight learned in stage 1 and the base class classification weight learned in stage 2 are inputted to the GAT to generate the reconstructed classification weights (prototype vectors).)
“receiving one or more second training datasets, each comprising second data samples of a set of one or more associated novel classes” ([0039], A novel class dataset is used to train the classifier. The new dataset contains data samples of a plurality of different classes.)
“” ([0035, 0037, 0040], The novel class classification weight (second output vector) is generated by the feature extractor based on the new dataset and the weights are provided to the GAT. Learning in the GAT is performed in an episodic format, which implies a memory to store the samples. The secondary reference will be combined to provide a more detailed disclosure of the use of an episodic memory.)
“retraining the classifier using the one or more second training datasets and the first training dataset by ” ([0039-0040, 0049], The classifier is trained for multiple sessions and are adapted to the dataset. The GAT learns the dependency between the basic dataset and the new dataset by meta learning.)
“determining a set of updated prototype vectors indicative of the first training dataset and the one or more second training datasets” ([0040], All the classification weights generated for the basic and novel dataset are provided to the GAT. The GAT generates the reconstructed classification weights (updated prototype vectors).)
Nishimaki does not explicitly disclose an implementation of “storing the one or more prototype vectors in the explicit memory”, “adding to the explicit memory one or more second output vectors”, “minimizing a distance between the one or more second output vectors and the one or more prototype vectors”, and “updating the explicit memory with the set of updated prototype vector”. However, Cheraghian discloses in the same field of endeavor:
“storing the one or more prototype vectors in the explicit memory” ([pg. 3, Section 3.1, par. 1; pg. 3-4, Section 3.2, par. 1; pg. 5, Algorithm 1], In the first task, there are training samples that are associated with the base class and the base task. In the following tasks, the training samples include novel classes. The feature representations for each class in the former tasks are saved into memory. From Algorithm 1, lines 4-11 shows the training of the model for the first task and in line 13, the memory is updated with the representations from the first task.)
“adding, to the explicit memory, one or more second output vectors indicative of the set of one or more associated novel classes, in response to providing the one or more second training datasets to the classifier” ([pg. 3, Section 3.1, par. 1; pg. 3-4, Section 3.2, par. 1; pg. 4, Figure 3; pg. 5, Algorithm 1], For tasks t > 1, the training samples include novel classes such as the example shown in Figure 3. The training samples are processed by the backbone to generate the feature representations. These representations are updated to the memory for each task.)
“retraining the classifier using the one or more second training datasets and the first training dataset by minimizing a distance between the one or more second output vectors and the one or more prototype vectors” ([pg. 4-5, Section 3.3, par. 1-4], The classifier is trained for a number of epochs. The representations of the classes of the first task are cluster using k-means clustering. The superclass label to novel classes are assign based on calculating the minimum Euclidean distance between the representations of the novel classes and the cluster centers, which are based on the classes of the first task. Each embedding is also trained using a cross-entropy loss based on the superclass labels.)
“updating the explicit memory with the set of updated prototype vectors” ([pg. 5, Section 3.4, par. 1, Algorithm 1, pg. 6, Figure 5], The classifier is trained with cross-entropy loss to generate the global feature representation. From Algorithm 1, line 28 shows the calculating of the protype vector by averaging all the training samples from each class and updating the memory. Training involves multiple epochs and it is an iterative process. The memory will be updated with new representations as new dataset are processed by the classifier.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “storing the one or more prototype vectors in the explicit memory”, “adding to the explicit memory one or more second output vectors”, “minimizing a distance between the one or more second output vectors and the one or more prototype vectors”, and “updating the explicit memory with the set of updated prototype vector” from Cheraghian into the teaching of Nishimaki. Doing so improve the catastrophic forgetting in class incremental learning by implementing a distillation algorithm (Cheraghian, abstract).
Regarding claim 2, Nishimaki in view of Cheraghian teaches:
“wherein the controller comprises a feature extractor and a classification head” ([Cheraghian, pg. 6, Section 4.1, par. 4], The model consists of a backbone network (feature extractor) and embedding modules with fully connected layers (classification head).)
Regarding claim 3, Nishimaki in view of Cheraghian teaches:
“wherein the feature extractor receives the first data sample and provides a first extracted feature vector, and wherein the classification head receives the first extracted feature vector and provides the set of first output vectors to indicate the one or more associated base classes of the first data sample” ([Cheraghian, pg. 5, Section 3.4, par. 1; pg. 6, Figure 5], The model receives a plurality of training samples that represent different classes and it is input into the backbone network to generate feature representations. The feature representations are fed into embedding modules to generate representations based on the superclass information.)
Regarding claim 4, Nishimaki in view of Cheraghian teaches:
“wherein the controller is a convolutional neural network controller comprising multiple nonlinear layers for the feature extractor and the classification head being an output layer of the convolutional neural network” ([Cheraghian, pg. 6, Section 4.1, par. 4; pg. 6, Figure 5], ResNet18 is employed for the backbone network and it is inherent ResNet18 comprises numerous nonlinear layer that uses ReLU activation function. The embedding modules consist of fully connected layers that outputs representations.)
Regarding claim 11, Nishimaki in view of Cheraghian discloses:
“receiving a query vector” ([Nishimaki, 0037, 0057], Each episode consists of a support set and a query set. The query set may be generated by the synthetic image generation unit.)
“performing a similarity search between the query vector and the one or more prototype vectors in the explicit memory for determining a class represented by the query vector” ([Nishimaki, 0037, 0057], The query samples are classified based on the support samples and the model is updated to minimize the loss in classification.)
Regarding claim 13, Nishimaki in view of Cheraghian discloses:
“wherein the retraining the classifier is a few shot learning performed upon the set of one or more associated novel classes” ([Nishimaki, 0008, 0039], The classifier is trained using few shot learning of a set of novel classes for multiple sessions.)
Regarding claim 14, Nishimaki in view of Cheraghian discloses:
“wherein the retraining the classifier comprises retraining a part of the classifier and freezing another pretrained part of the classifier” ([Nishimaki, 0034], The parameter of the feature extractor of the pre-trained backbone CNN is fixed.)
Regarding claim 15, Nishimaki in view of Cheraghian discloses:
“wherein the prototype vector comprises a set of elements each indicating a probability that a respective class of the one or more associated base classes is a class of the one or more prototype vectors” ([Nishimaki, 0034-0036], The base classification weight (probability) indicates the average feature amount of the data sample of the basic dataset having a plurality of the base class. The base class classification weight are processed by the GAT to generate the reconstructed classification weights (prototype vector).)
Regarding claim 16:
Claim 16 recites an article of manufacture that performs the same process as described in Claim 1. Therefore claim 16 is rejected under the same reasons mention for claim 1. The additional elements of claim 16 is addressed below by Nishimaki:
“A computer program product comprising a computer-readable storage medium having computer-readable program code embodied therewith, wherein the computer-readable program code when called by a processor causes the processor to” ([0088], The machine learning apparatus may be implemented by devices like a CPU and a memory.)
Regarding claim 17:
Claim 17 recites a system that performs the same process as described in Claim 1. Therefore claim 17 is rejected under the same reasons mention for claim 1. The additional elements of claim 17 is addressed below Nishimaki:
“A computer system for continual training of a classifier, the classifier comprising a controller and an explicit memory, the computer system comprising a processor and a computer-readable storage medium having computer-readable program code embodied therewith, wherein the computer-readable program code when called by the processor causes the processor to” ([0035, 0088], The apparatus consists of a feature extraction unit (controller) and a graph attention network that learns in an episodic format, which implies a memory that stores intermediate outputs. The machine learning apparatus may be implemented by devices like a CPU and a memory.)
Claims 5 is rejected under 35 U.S.C. 103 as being unpatentable over Nishimaki (US20250200397A1) in view of Cheraghian, “Semantic-aware Knowledge Distillation for Few-Shot Class-Incremental Learning” and Han, “Continual Relation Learning via Episodic Memory Activation and Reconsolidation”.
Regarding claim 5, Nishimaki in view of Cheraghian teaches:
“inferring the feature extractor using the one or more second training datasets and storing, in an , second extracted feature vectors of each of the second data samples of the one or more second training datasets” ([Cheraghian, pg. 5, Section 3.4, par. 1; pg. 6, Section 4.1, par. 1; pg. 5, Algorithm 1], The experiments are conducted for 3 datasets and each dataset contains a number of base and novel classes. The backbone network generates the global feature representation and the prototype vector is updated in memory during training.)
“training the classification head using all stored extracted feature vectors” ([Cheraghian, pg. 5, Section 3.4, par. 1; pg. 5, Algorithm 1], The global feature representation is fed into a number of embedding modules for training the modules.)
“wherein the training of the classification head is performed by minimizing a distance between the one or more second output vectors and the set of updated prototype vectors” ([Cheraghian, pg. 4, Section 3.3, par. 1-2], The embedding modules are trained by using a cross-entropy loss based on the superclass labels obtained in the previous stage. The minimum Euclidean distance is calculated between the semantic vector of novel class and the cluster centers, which are determined based on the classes of the first task.)
Nishimaki in view of Cheraghian does not explicitly disclose an implementation of “storing in an activation memory”. However, Han discloses in the same field of endeavor:
“... storing, in an activation memory, second extracted feature vectors of each of the second data samples of the one or more second training datasets” ([pg. 5, Section 3.5, par. 2-3; pg. 6, Section 4.1, par. 1-4], The experiments are conducted for 3 datasets. Memory replay and activation is applied to continually learn new relations and remember old relations. The activation set are stored and retrieved from memory during memory replay and activation for multiple iterations.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “storing in an activation memory” from Han into the teaching of Nishimaki in view of Cheraghian. Doing so improve the continual learning models with the implementation of a episodic memory activation and reconsolidation for continual relation learning (Han, abstract).
Claims 6-7 are rejected under 35 U.S.C. 103 as being unpatentable over Nishimaki (US20250200397A1) in view of Cheraghian, “Semantic-aware Knowledge Distillation for Few-Shot Class-Incremental Learning”, Han, “Continual Relation Learning via Episodic Memory Activation and Reconsolidation” and Ren, “Incremental Few-Shot Learning with Attention Attractor Networks”.
Regarding claim 6, Nishimaki in view of Cheraghian and Han teaches:
“wherein the multiple nonlinear layers with parameter set θ1 and the classification head with parameter set θ2 produce the one or more second output vectors and predict a probability p over a combined set of the set of one or more associated base classes and the set of one or more associated novel classes, and wherein the distance between the one or more second output vectors and the set of updated prototype vectors is minimized by using
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
, where ki* is the set of updated prototype vectors in association with an ith class, and ai is an associated extracted feature vector of the classification head for the ith class” ([Cheraghian, pg. 3-4, Section 3.2, par. 1; pg. 4-5, Section 3.3, par. 2-3; pg. 6, Figure 5], The backbone network (multiple nonlinear layers) consists of a set of training parameters and the embedding modules (classification head) are trained on a set of parameters to produce weighted average as the embedding representation. The minimum Euclidean distance is determined the semantic vector of novel class and the cluster centers. Nishimaki in view of Cheraghian and Han does not disclose the optimization process using a gradient descent algorithm, which is what is being described by the equation for updating the parameters of a classifier.)
Nishimaki in view of Cheraghian and Han does not explicitly disclose an implementation of “
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
, where
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
”. However, Ren discloses in the same field of endeavor:
“... the distance between the one or more second output vectors and the set of updated prototype vectors is minimized by using:
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
, where
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
,” ([pg. 5-6, Section 3.3, par. 1-2; pg. 2, Figure 1], A gradient descent process is applied to update the parameters of the model. The model learns base class weights using a backbone network and the attention attractor network learns the novel classes.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “
θ
2
(
t
+
1
)
=
θ
2
(
t
)
-
β
∂
L
F
θ
2
(
t
)
,
K
*
,
A
)
∂
θ
2
(
t
)
, where
L
F
=
-
∑
i
=
1
c
S
h
k
i
,
*
,
g
θ
2
t
(
a
i
)
” from Ren into the teaching of Nishimaki in view of Cheraghian and Han. Doing so improve the performance of a classification model by implementing the technique of recurrent back-propagation to back-propagate through the optimization process and facilitate the learning of these parameters (Ren, abstract).
Regarding claim 7, Nishimaki in view of Cheraghian and Han teaches:
“wherein the one or more prototype vectors p* •* is defined for the ith class as follows: p*i=
g
θ
2
T
(
a
i
)
” ([Cheraghian, pg. 3-4, Section 3.2, par. 1], The mapping module performs a function on the feature representation to project it into the semantic domain. yi = M(gi) describes the function to obtain the projected feature.)
Claims 8-10, 12, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Nishimaki (US20250200397A1) in view of Cheraghian, “Semantic-aware Knowledge Distillation for Few-Shot Class-Incremental Learning” and Karunaratne, “Robust High-dimensional Memory-augmented Neural Networks”.
Regarding claim 8, Nishimaki in view of Cheraghian teaches:
“prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated ” ([Cheraghian, pg. 5, Section 3.4, par. 1; pg. 5, Algorithm 1], A prototype vector is calculate for each class by averaging of all training samples from each class and the memory is updated with the newly determined prototype vector.)
Nishimaki in view of Cheraghian does not explicitly disclose an implementation of “prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated quasi-orthogonal prototype vectors and resultantly updating the explicit memory with the quasi-orthogonal prototype vectors”. However, Karunaratne discloses in the same field of endeavor:
“prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated quasi-orthogonal prototype vectors and resultantly updating the explicit memory with the quasi-orthogonal prototype vectors” ([pg. 2-3, Section II.A, par. 2; pg. 3-4, Section II.B, par. 1-4], A attention function finds quasi-orthogonal vectors in the HD feature space for complex image inputs. The quasi-orthogonal vectors are stored in associative memory.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated quasi-orthogonal prototype vectors and resultantly updating the explicit memory with the quasi-orthogonal prototype vectors” from Karunaratne into the teaching of Nishimaki in view of Cheraghian. Doing so improve the training of neural networks by implementing computation of high-dimensional vectors (Karunaratne, abstract).
Regarding claim 9, Nishimaki in view of Cheraghian does not explicitly disclose the elements of claim 9. However, Karunaratne discloses in the same field of endeavor:
“wherein determining the quasi-orthogonal prototype vectors comprises backpropagation using a loss function as follows:
PNG
media_image2.png
194
620
media_image2.png
Greyscale
--
where
k
i
(
u
)
is an quasi-orthogonal prototype vector obtained in iteration number u for the ith class, and pi is a prototype vector stored in the explicit memory in association with the ith class, and sh denotes a soft hamming distance, where the initial quasi-orthogonalized prototypes are given by
k
i
(
0
)
=
p
i
” ([pg. 5-6, Section II.D, par. 1-2; pg. 29, Section 1.c, par. 1-4; pg. 30-32, Section 2], A logarithm loss is computed for every query based on the output of the previous step and the one-hot labels. The controller is trained with the logarithmic loss through backpropagation. The vectors shown in Equations 5-10 (pg. 32) are quasi-orthogonal vectors in high dimensional space. The high dimensional vectors are converted to binary vectors and the dot products are computed to determine the similarities (soft hamming distance).)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “prior to retraining the classifier, modifying the one or more prototype vectors thereby determining associated quasi-orthogonal prototype vectors and resultantly updating the explicit memory with the quasi-orthogonal prototype vectors” from Karunaratne into the teaching of Nishimaki in view of Cheraghian. Doing so improve the training of neural networks by implementing computation of high-dimensional vectors (Karunaratne, abstract).
Regarding claim 10, Nishimaki in view of Cheraghian and Karunaratne discloses:
“providing an activation memory for accumulating the extracted feature vectors of the feature extractor” ([Karunaratne, pg. 6, Section II.D, par. 2; pg. 29, Section 1.a; pg. 18, Figure 4], A controller maps training data to the feature space and stores it into key memory. Figure 4 shows a plurality of memory components arranged in an array. A separate value memory is also present to store the one-hot support labels from the controller. Multiple memory is disclose for storing various data.)
“wherein the classification head is configured to receive an input extracted feature vector from the activation memory” ([Karunaratne, pg. 9-10, Section “The CNN as a controller for the MANN architecture”], A controller contains a fully connected layer. During learning and inference, the outputs are stored into the key memory and may be retrieved.)
Regarding claim 12, Nishimaki in view of Cheraghian and Karunaratne discloses:
“providing an in-memory computing core comprising a crossbar array structure comprising row lines and column lines and resistive memory elements coupled between the row lines and the column lines at junctions formed by the row and column lines” ([Karunaratne, pg. 6, Section II.D, par. 2; pg. 7, Section III, par. 3; pg. 18, Figure 4], Figure 4 shows the crossbar array structure consisting of multiple memories. The key memory may also be based on resistive random access memory.)
“programming the resistive memory elements of each column line to represent values of the one or more prototype vectors” ([Karunaratne, pg. 30, Section 2, par. 1; pg. 20, Supplementary Figure 1], The key memory is programmed for the few-shot classification problem.)
“inputting, to the crossbar array, the query vector for performing the similarity search” ([Karunaratne, pg. 30, Section 2, par. 1; pg. 18, Figure 4], The query samples are stored in the crossbar array.)
Regarding claim 18, Nishimaki in view of Cheraghian and Karunaratne discloses:
“a crossbar array structure comprising row lines and column lines and resistive memory elements coupled between the row lines and the column lines at junctions formed by the row and column lines, the resistive memory elements of each column line representing values of a respective prototype vector, the crossbar array being configured to receive elements of a query vector through the row lines respectively and to perform a vector matrix multiplication at the crossbar array for computing a similarity of the query vector with the one or more prototype vectors, thereby determining a class of the query vector” ([Karunaratne, pg. 5-6, Section II.D, par. 1-2; pg. 7, Section III, par. 3; pg. 30, Section 2, par. 1; pg. 18, Figure 4], Figure 4 shows the crossbar array structure consisting of multiple memories. The key memory may also be based on resistive random access memory. The key memory is programmed for the few-shot classification problem and stores the outputs from the controller. The query samples are stored in the crossbar array. Pairwise cosine similarity is performed between query vectors and the support vectors to determine the classification.)
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GARY MAC whose telephone number is (703)756-1517. The examiner can normally be reached Monday - Friday 8:00 AM - 5:00 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 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.
/GARY MAC/Examiner, Art Unit 2127 /ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights