DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment/Status of Claims
Claims 1, 4, and 5 were amended.
Claims 9 and 10 are new.
Claims 1-5 and 7-10 are pending and examined herein.
Claims 1-5 and 7-10 are rejected under 35 U.S.C. 101.
Claims 1-5 and 7-10 are rejected under 35 U.S.C. 103.
Response to Arguments
Applicant's arguments filed 2/17/2026 have been fully considered but they are not persuasive.
Applicant argues, see pages 5-6, "Learning a parameter... by maximizing an expected value... using a gradient estimation method. Maximizing the expected value of an Adjusted Rand Index (ARI) across high-dimensional feature spaces generated by a deep neural network is not a task that can be practically performed in the human mind. It is a sophisticated machine learning technique that requires the high-speed processing capabilities of a processor and memory to perform the millions of iterative calculations necessary for gradient-based optimization."
Examiner agrees that the limitation challenged is not a mental process. However, MPEP 2106(a)(2)(I)(C) states "A claim that recites a mathematical calculation, when the claim is given its broadest reasonable interpretation in light of the specification, will be considered as falling within the ‘mathematical concepts’ grouping. A mathematical calculation is a mathematical operation (such as multiplication) or an act of calculating using mathematical methods to determine a variable or number, e.g., performing an arithmetic operation such as exponentiation. There is no particular word or set of words that indicates a claim recites a mathematical calculation. That is, a claim does not have to recite the word ‘calculating’ in order to be considered a mathematical calculation. For example, a step of ‘determining’ a variable or number using mathematical methods or ‘performing’ a mathematical operation may also be considered mathematical calculations when the broadest reasonable interpretation of the claim in light of the specification encompasses a mathematical calculation." MPEP 2106(a)(2)(I)(C) lists examples of mathematical calculations including "v. using an algorithm for determining the optimal number of visits by a business representative to a client, In re Maucorps, 609 F.2d 481, 482, 203 USPQ 812, 813 (CCPA 1979)." Therefore, "Learning a parameter by maximizing an expected value using a gradient estimation method,” recites a mathematical calculation, which falls under the abstract idea grouping of mathematical concepts.
Applicant further argues, see pages 6-7 "Claim 1 achieves this by reciting a specific technological process that improves clustering performance for complex data distributions-a task that conventional generic computers and human mental processes fail to perform with high accuracy."
Particularly, Applicant contends that the specification states that "Standard clustering methods, such as infinite Gaussian mixture models, often exhibit deteriorated performance when clusters cannot be represented by a standard Gaussian distribution." Additionally, Applicant states that the specific technical solution is "Claim 1 does not merely "calculate" data; it utilizes a predetermined neural network to transform raw data into a plurality of representation data items (feature vectors) specifically optimized for clustering," citing paragraphs [0006], [0017]-[0018], [0045]. Applicant further presents evidence that "The Specification provides experimental results showing that the "Proposed Method" (optimizing based on the evaluation scale) achieves an Adjusted Rand Index (ARI) of 0.912, significantly outperforming existing technologies such as standard GMM (0.882) and AE + GMM (0.866)."
MPEP 2106.05(a) further states "It is important to note, the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements. See the discussion of Diamond v. Diehr, 450 U.S. 175, 187 and 191-92, 209 USPQ 1, 10 (1981)) in subsection II, below. In addition, the improvement can be provided by the additional element(s) in combination with the recited judicial exception. See MPEP § 2106.04(d) (discussing Finjan, Inc. v. Blue Coat Sys., Inc., 879 F.3d 1299, 1303-04, 125 USPQ2d 1282, 1285-87 (Fed. Cir. 2018)). Thus, it is important for examiners to analyze the claim as a whole when determining whether the claim provides an improvement to the functioning of computers or an improvement to other technology or technical field."
As Applicant states that the technical solution is “utiliz[ing] a predetermined neural network to transform raw data into a plurality of representation data items (feature vectors) specifically optimized for clustering”, the limitations that the proposed improvement are evident in appear to be “converting each of the plurality of items of data … to generate a plurality of items of representation data; clustering the plurality of items of representation data; calculating a predetermined evaluation scale that evaluates a quality of clustering division, including an Adjusted Rand Index, indicating performance of the clustering, based on the clustering result and the plurality of labels; and learning a parameter of the neural network by maximizing an expected value of the evaluation scale using a gradient estimation method.” These limitations are all analyzed as judicial exceptions (see 35 U.S.C. 101 claim rejections below), and cannot provide an improvement to the functioning of computers.
Applicant’s arguments, see pages 8-11, filed 02/17/2026, with respect to the rejection(s) of claim(s) 1-5 and 7-8 under 35 U.S.C. 102 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Allen (“Infinite Mixture Prototypes for Few-Shot Learning”, 2019) and Fritsch (“Improved Criteria for Clustering Based on the Posterior Similarity Matrix”, 2009).
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 1-5 and 7-10 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
MPEP § 2109(III) sets out steps for evaluating whether a claim is drawn to patent-eligible subject
matter. The analysis of claims 1-5 and 7-8, in accordance with these steps, follows.
Step 1 Analysis:
Step 1 is to determine whether the claim is directed to a statutory category (process, machine,
manufacture, or composition of matter. Claims 1-4 and 9-10 are directed to processes, claim 5 is directed to a machine, and claims 7-8 are directed to articles of manufacture. All claims are directed to statutory categories and analysis continues.
Step 2A Prong One, Step 2A Prong Two, and Step 2B Analysis:
Step 2A Prong One asks if the claim recites a judicial exception (abstract idea, law of nature, or natural phenomenon). If the claim recites a judicial exception, analysis proceeds to Step 2A Prong Two, which asks if the claim recites additional elements that integrate the abstract idea into a practical application. If the claim does not integrate the judicial exception, analysis proceeds to Step 2B, which asks if the claim amounts to significantly more than the judicial exception. If the claim does not amount to significantly more than the judicial exception, the claim is not eligible subject matter under 35 U.S.C. 101.
None of the claims represent an improvement to technology.
Regarding claim 1, the following claim elements are abstract ideas:
converting each of the plurality of items of data … to generate a plurality of items of representation data; (Converting data into representations of that data can be practically performed in the human mind. This is a mental process.)
clustering the plurality of items of representation data; (Clustering the representation data can be practically performed in the human mind. This is a mental process of evaluation.)
calculating a predetermined evaluation scale that evaluates a quality of clustering division, including an Adjusted Rand Index, indicating performance of the clustering, based on the clustering result and the plurality of labels; and (Calculating a performance measure, including an Adjusted Rand Index, is a mathematical calculation, which is a mathematical concept.)
learning a parameter of the neural network by maximizing an expected value of the evaluation scale using a gradient estimation method. (Maximizing a value using a gradient estimation method is a mathematical calculation, which is a mathematical concept.)
The following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
A learning method, executed by a computer including a memory and a processor, the method comprising: (This limitation recites generic computer parts and processes; this amounts to mere instructions to apply an exception.)
inputting a plurality of items of data, and a plurality of labels representing clusters to which the plurality of items of data belong; (Data input is the generic computer process of transmitting data, which amounts to mere instructions to apply an exception.)
by a predetermined neural network, (This recites a generic neural network; this amounts to mere instructions to apply an exception.)
Regarding claim 2, the rejection of claim 1 is incorporated herein. The following is an abstract idea:
wherein the converting converts each of the plurality of items of data and data representing a representation of a predetermined target task to generate the plurality of items of representation data. (One could convert data to generate representations practically in the human mind. This is an mental process.)
The following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
by the neural network (This recites a generic neural network; this amounts to mere instructions to apply an exception.)
Regarding claim 3, the rejection of claim 1 is incorporated herein. The following are abstract ideas:
wherein the clustering performs clustering by estimating a contribution rate indicating a probability that each of the plurality of items of representation data belongs to each of the plurality of clusters, and (Estimating a contribution rate can be practically performed in the human mind. This is a mental process.)
the calculating calculates the evaluation scale by using the contribution rate as the clustering result. (This is a mathematical calculation, which is a mathematical concept.)
Regarding claim 4, the rejection of claim 1 is incorporated herein. The following are abstract ideas:
converting each of the plurality of items of data … to generate a plurality of items of representation data; and (Converting data into representations of that data can be practically performed in the human mind. This is a mental process.
clustering the plurality of items of representation data. (Clustering the representation data can be practically performed in the human mind. This is a mental process of evaluation.)
The following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
A clustering method, executed by a computer including a memory and a processor, the method comprising: (This limitation recites generic computer parts and processes; this amounts to mere instructions to apply an exception.)
inputting a plurality of items of data; (Data input is the generic computer process of transmitting data, which amounts to mere instructions to apply an exception.)
by a predetermined neural network in which a parameter trained in advance is set, said predetermined neural network being trained in advance by the learning method according to claim 1. (This recites a generic neural network and generic training; this amounts to mere instructions to apply an exception. See above explanation of claim 1.)
Regarding claim 5, the following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
A learning apparatus comprising: (This recites a generic machine learning apparatus; this amounts to mere instructions to apply an exception.)
a memory and a processor configured to (This limitation recites generic computer parts and processes; this amounts to mere instructions to apply an exception.)
The remainder of claim 5 recites substantially similar subject matter to claim 1 and is rejected with the same rationale, mutatis mutandis.
Regarding claim 7, the following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to execute the learning method as set forth in claim 1. (This limitation recites generic computer parts and processes; this amounts to mere instructions to apply an exception. See claim 1 for the remainder of the analysis.)
Regarding claim 8, the following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to execute the clustering method as set forth in claim 4. (This limitation recites generic computer parts and processes; this amounts to mere instructions to apply an exception. See claim 1 for the remainder of the analysis.)
Regarding claim 9, the rejection of claim 1 is incorporated herein. The following are abstract ideas:
the clustering includes estimating a contribution rate indicating a probability that each of the plurality of items of representation data belongs to each of a plurality of clusters; (Estimating a contribution rate can be practically performed in the human mind. This is a mental process.)
the calculating includes calculating the Adjusted Rand Index as the predetermined evaluation scale based on the estimated contribution rate and the plurality of labels; and (Calculating the Adjusted Rand Index is a mathematical calculation, which is a mathematical concept.)
the learning includes learning the parameter of the neural network by maximizing an expected value of the Adjusted Rand Index using the gradient estimation method. (Maximizing an expected value using a gradient estimation method is performing a specific algorithm, which amounts to mathematical calculations. This is a mathematical concept.)
Regarding claim 10, the rejection of claim 1 is incorporated herein. The following are abstract ideas:
the clustering is performed by clustering the plurality of items of representation data by estimating a probability that each item of the representation data belongs to each cluster using a variational Bayesian method; (Using a variational Bayesian method to estimate probabilities to cluster data is performing a specific algorithm, which amounts to mathematical calculations. This is a mental process.)
the calculating is performed by calculating a non-differentiable clustering evaluation scale including an Adjusted Rand Index (ARI) based on the estimated probability and the plurality of labels; and (Calculating the Adjusted Rand Index is a mathematical calculation, which is a mathematical concept.)
the learning is performed by learning a parameter of the neural network by maximizing an expected value of the non-differentiable clustering evaluation scale using a gradient estimation method to optimize the neural network for generating the representation data that forms a cluster structure consistent with the plurality of labels. (Maximizing an expected value using a gradient estimation method is performing a specific algorithm, which amounts to mathematical calculations. This is a mathematical concept.)
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-5 and 7-9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Allen (“Infinite Mixture Prototypes for Few-Shot Learning”, 2019) and Fritsch (“Improved Criteria for Clustering Based on the Posterior Similarity Matrix”, 2009).
Regarding claim 1, Allen teaches
A learning method, executed by a computer including a memory and a processor, the method comprising: (Page 1 states "We propose infinite mixture prototypes to adaptively represent both simple and complex data distributions for few-shot learning." One of ordinary skill in the art would realize that a machine learning method would be implemented on a computer, which necessarily includes a memory and a processor to execute the machine learning method.)
inputting a plurality of items of data, and a plurality of labels representing clusters to which the plurality of items of data belong; (Page 2 states "In few-shot classification we are given a support set
S
=
x
1
,
y
1
,
…
,
(
x
K
,
y
K
}
of
K
labeled points and a query set
Q
=
{
x
1
'
,
y
1
'
,
…
,
x
K
'
,
y
K
'
of
K
'
labeled points where each
x
i
,
x
i
'
∈
R
D
is a
D
-dimensional feature vector and
y
i
,
y
i
'
∈
{
1
,
…
,
N
}
is the corresponding label." The support set points and labels are interpreted as the plurality of items of data and the plurality of labels. Page 4, Algorithm 1 states “Init. each cluster
μ
c
with label
l
c
”. Therefore, the labels represent the clusters to which the plurality of items of data belong.)
converting each of the plurality of items of data by a predetermined neural network, to generate a plurality of items of representation data; (Page 2 states "For nonparametric representation learning methods, the model parameters are for the embedding function
h
ϕ
:
R
D
→
R
M
that map an input point
x
into a feature. The embedding of point
x
is the
M
-dimensional feature vector from the embedding function. In deep models the parameters
ϕ
are the weights of a deep network, and the embedding is the output of the last layer of this network." Page 10 states "For all few-shot experiments, we use the same base architecture as prototypical networks for the embedding network." Page 4, Algorithm 1 states “Init. each cluster
μ
c
with label
l
c
and
σ
c
=
σ
l
as class-wise means of the supports”. Page 3 states “Prototypical networks (Snell et al., 2017) form prototypes as the mean of the embedded support points in each class:
μ
n
=
1
|
S
n
|
∑
x
i
,
y
i
∈
S
n
h
ϕ
x
i
.” Thus, a neural network is used to convert the items of data into embeddings, interpreted as the representation data.)
clustering the plurality of items of representation data; (Page 4, Algorithm 1 states “Assign supports to clusters”. Therefore, the items are clustered.)
calculating a predetermined evaluation scale indicating performance of the clustering. based on the clustering result and the plurality of labels; and (Page 4 states "We optimize all models with the cross-entropy loss. For the multi-modal methods (nearest neighbors and IMP), we mask the loss to only include the closest neighbor/cluster for each class, in the same manner as inference. That is, for a class
n
, we find the most likely cluster
c
n
*
←
argmax
c
:
l
c
=
n
log
p
(
h
ϕ
(
x
)
|
μ
c
,
σ
c
)
and then take the loss over the queries in the class (
Q
n
):". The loss is interpreted as the predetermined evaluation scale which, as it includes a loss of the clustering, indicates the performance of the clustering.)
learning a parameter of the neural network … using a gradient estimation method. (Page 4 states "IMP optimizes the embedding parameters
ϕ
and cluster variances
σ
by stochastic gradient descent across episodes." As above, the models are optimized through the loss function, interpreted as the evaluation scale.)
Allen does not appear to explicitly teach
[an evaluation scale] that evaluates a quality of clustering division, including an Adjusted Rand Index
maximizing an expected value of the evaluation scale
However, Fritsch—directed to analogous art—teaches
[an evaluation scale] that evaluates a quality of clustering division, including an Adjusted Rand Index (Page 372 states "The adjusted Rand has the usual form of an index corrected for chance:
Index
-
Expected Index
Maximum Index
-
Expected Index
. It has a maximum value of 1 and its value is 0 if the Rand index equals its expected value." The adjusted Rand Index is interpreted as the evaluation scale that evaluates a quality of clustering division.)
maximizing an expected value of the evaluation scale (Page 373 states "The adjusted Rand index with the posterior expected clustering
A
R
c
*
,
E
c
y
is given by the expression [Expression (13)]." Page 374 states "Expression (13) requires the computation of the posterior similarity matrix, but can then be evaluated a lot faster than (12), which is advantageous if the criterion needs to be calculated for many different
c
*
. We also found (13) to be more amenable to a theoretical study. Expression (12) on the other hand does not require the computation of the posterior similarity matrix, which can be preferable for large n, where this matrix gets too large to be stored. And one can argue that maximizing
E
(
A
R
(
c
*
;
c
)
|
y
)
is a more standard approach than maximizing
A
R
c
*
,
E
c
y
. Practically, however, we found in our applications that maximization of either criterion leads to nearly identical results. For simplicity we will in the following refer to both criteria as PEAR for Posterior Expected Adjusted Rand." Maximizing PEAR/
A
R
c
*
,
E
c
y
/Expression (13) is interpreted as the maximization of the expected value.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Allen with the teachings of Fritsch because, as Fritsch states on page 367, "We propose new criteria for estimating a clustering, which are based on the posterior expected adjusted Rand index. The criteria are shown to possess a shrinkage property and outperform Binder's loss in a simulation study and in an application to gene expression data. They also perform favorably compared to other clustering procedures." Note that, with the differentiable expression taught by Fritsch, see Expression (13) on page 373, which is further simplified in Expression (14) on page 374, one of ordinary skill in the art would be motivated to use the gradient estimation method taught by Allen.
Regarding claim 2, the rejection of claim 1 is incorporated herein. Allen teaches
wherein the converting converts each of the plurality of items of data and data representing a representation of a predetermined target task by the neural network to generate the plurality of items of representation data. (Page 2 states "Few-shot classification is commonly learned by constructing few-shot tasks from a large dataset and optimizing the model parameters on these tasks. Each task, comprised of the support and query sets, is called an episode. Episodes are drawn from a dataset by randomly sampling a subset of classes, sampling points from these classes, and then partitioning the points into supports and queries." Therefore, each task is comprised of plurality of data points which are interpreted as the plurality of items of data. Page 2 states "For nonparametric representation learning methods, the model parameters are for the embedding function
h
ϕ
:
R
D
→
R
M
that map an input point
x
into a feature. The embedding of point
x
is the
M
-dimensional feature vector from the embedding function. In deep models the parameters
ϕ
are the weights of a deep network, and the embedding is the output of the last layer of this network." Page 10 states "For all few-shot experiments, we use the same base architecture as prototypical networks for the embedding network." Page 4, Algorithm 1 states “Init. each cluster
μ
c
with label
l
c
and
σ
c
=
σ
l
as class-wise means of the supports”. Page 3 states “Prototypical networks (Snell et al., 2017) form prototypes as the mean of the embedded support points in each class:
μ
n
=
1
|
S
n
|
∑
x
i
,
y
i
∈
S
n
h
ϕ
x
i
.” Thus, a neural network is used to convert the items of data into embeddings, interpreted as the representation data. As the items of data in each task are converted, the task representation are converted.)
Regarding claim 3, the rejection of claim 1 is incorporated herein. Allen teaches
wherein the clustering performs clustering by estimating a contribution rate indicating a probability that each of the plurality of items of representation data belongs to each of the plurality of clusters, and (Page 4 states "Finally, cluster means
μ
c
are computed by the weighted mean of their members. Since each class can have multiple clusters, we classify a query point
x
'
by the softmax over distances to the closest cluster in each class
n
:
p
ϕ
y
'
=
n
x
=
exp
-
d
h
ϕ
x
'
,
μ
c
n
*
∑
n
'
exp
(
-
d
h
ϕ
,
μ
c
n
'
*
)
)
with
c
n
*
=
arg
min
c
:
l
c
=
n
d
(
h
ϕ
x
'
,
μ
c
)
indexing the clusters, where each cluster has label
l
c
." One of ordinary skill in the art would realize that
p
ϕ
y
'
=
n
x
are the probabilities that each piece of data
x
, which are representation data because they are embedded by the function
h
ϕ
in the equation, belong to each cluster with label
y
'
=
n
. This is further supported by the use of softmax, which as one of ordinary skill in the art would realize, converts numbers into a probability distribution.
p
ϕ
y
'
=
n
x
is interpreted as the contribution rate.)
the calculating calculates the evaluation scale by using the contribution rate as the clustering result. (Page 4 states “We optimize all models with the cross-entropy loss. For the multi-modal methods (nearest neighbors and IMP), we mask the loss to only include the closest neighbor/cluster for each class, in the same manner as inference. That is, for a class
n
, we find the most likely cluster
c
n
*
←
argmax
c
:
l
c
=
n
log
p
(
h
ϕ
(
x
)
|
μ
c
,
σ
c
)
and then take the loss over the queries in the class (
Q
n
):” As the probabilities are computed using the contribution rate function to find the maximum probability to use for the loss, the calculation of the loss/evaluation scale uses the contribution rate as the clustering result.)
Regarding claim 4, the rejection of claim 1 is incorporated herein. Allen teaches
A clustering method, executed by a computer including a memory and a processor, the method comprising: (Page 1 states "We therefore propose Infinite Mixture Prototypes (IMP) to represent a class as a set of clusters, with the number of clusters determined directly from the data." One of ordinary skill in the art would realize that a machine learning method would be implemented on a computer, which necessarily includes a memory and a processor to execute the machine learning method.)
inputting a plurality of items of data; (Page 2 states "In few-shot classification we are given a support set
S
=
x
1
,
y
1
,
…
,
(
x
K
,
y
K
}
of
K
labeled points and a query set
Q
=
{
x
1
'
,
y
1
'
,
…
,
x
K
'
,
y
K
'
of
K
'
labeled points where each
x
i
,
x
i
'
∈
R
D
is a
D
-dimensional feature vector and
y
i
,
y
i
'
∈
{
1
,
…
,
N
}
is the corresponding label." Page 2 further states "The support set is for learning while the query set is for inference: the few-shot classification problem is to recognize the class of the queries given the labeled supports." The query set is interpreted as the plurality of items of data. )
converting each of the plurality of items of data by a predetermined neural network in which a parameter trained in advance is set, to generate a plurality of items of representation data, said predetermined neural network being trained in advance by the learning method according to claim 1; and (Page 2 states "Few-shot classification is commonly learned by constructing few-shot tasks from a large dataset and optimizing the model parameters on these tasks. Each task, comprised of the support and query sets, is called an episode. Episodes are drawn from a dataset by randomly sampling a subset of classes, sampling points from these classes, and then partitioning the points into supports and queries." Therefore, each task is comprised of plurality of data points which are interpreted as the plurality of items of data. Page 2 states "For nonparametric representation learning methods, the model parameters are for the embedding function
h
ϕ
:
R
D
→
R
M
that map an input point
x
into a feature. The embedding of point
x
is the
M
-dimensional feature vector from the embedding function. In deep models the parameters
ϕ
are the weights of a deep network, and the embedding is the output of the last layer of this network." Page 10 states "For all few-shot experiments, we use the same base architecture as prototypical networks for the embedding network." Page 4, Algorithm 1, which is for both the support set and the query set, states “Init. each cluster
μ
c
with label
l
c
and
σ
c
=
σ
l
as class-wise means of the supports”. Page 3 states “Prototypical networks (Snell et al., 2017) form prototypes as the mean of the embedded support points in each class:
μ
n
=
1
|
S
n
|
∑
x
i
,
y
i
∈
S
n
h
ϕ
x
i
.” Thus, a neural network is used to convert the items of data into embeddings, interpreted as the representation data. As the items of data in each task are converted, the task representation are converted.)
a clustering procedure of clustering the plurality of items of representation data. (Page 4, Algorithm 1 states “Assign supports to clusters”. Therefore, the items are clustered.)
Regarding claim 5, Allen teaches
A learning apparatus comprising: (Page 1 states "We propose infinite mixture prototypes to adaptively represent both simple and complex data distributions for few-shot learning." The computer that is used to perform the machine learning method is interpreted as the learning apparatus.)
a memory and a processor configured to (One of ordinary skill in the art would realize that a machine learning method would be implemented on a computer, which necessarily includes a memory and a processor to execute the machine learning method.)
The remainder of claim 5 recites substantially similar subject matter to claim 1 and is rejected with the same rationale, mutatis mutandis.
Regarding claim 7, the rejection of claim 1 is incorporated herein. Allen teaches
A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to execute the learning method as set forth in claim 1. (Page 1 states "We propose infinite mixture prototypes to adaptively represent both simple and complex data distributions for few-shot learning." One of ordinary skill in the art would realize that a machine learning method would be implemented on a computer, which necessarily includes a non-transitory computer-readable recording medium having computer-readable instructions for the method and a processor to execute the machine learning method.)
Regarding claim 8, the rejection of claim 4 is incorporated herein. Allen teaches
A non-transitory computer-readable recording medium having computer-readable instructions stored thereon, which when executed, cause a computer to execute the clustering method as set forth in claim 4. (Page 1 states "We therefore propose Infinite Mixture Prototypes (IMP) to represent a class as a set of clusters, with the number of clusters determined directly from the data." One of ordinary skill in the art would realize that a machine learning method would be implemented on a computer, which necessarily includes a non-transitory computer-readable recording medium having computer-readable instructions for the method and a processor to execute the machine learning method.)
Regarding claim 9, the rejection of claim 1 is incorporated herein. Allen teaches
the learning includes learning the parameter of the neural network … using the gradient estimation method. (Page 4 states "IMP optimizes the embedding parameters
ϕ
and cluster variances
σ
by stochastic gradient descent across episodes." As above, the models are optimized through the loss function, interpreted as the evaluation scale.)
Allen does not appear to explicitly teach
the clustering includes estimating a contribution rate indicating a probability that each of the plurality of items of representation data belongs to each of a plurality of clusters;
the calculating includes calculating the Adjusted Rand Index as the predetermined evaluation scale based on the estimated contribution rate and the plurality of labels; and
by maximizing an expected value of the Adjusted Rand Index
However, Fritsch—directed to analogous art—teaches
the clustering includes estimating a contribution rate indicating a probability that each of the plurality of items of representation data belongs to each of a plurality of clusters; (Page 369 states "If
K
varies between the clusterings a possible solution is to choose a clustering based on the posterior similarity matrix
P
c
i
=
c
j
y
)
, an
n
×
n
matrix that contains the pairwise probabilities that two observations belong to the same cluster. This approach is taken, for example, in the Bayesian cluster models for microarray data by Dahl (2006) and Medvedovic, Yeung, and Bumgarner (2004)."
P
c
i
=
c
j
y
)
is interpreted as the contribution rate, as it indicates the probabilities of each pair of items of representation data
i
,
j
, which includes each of the items of representation data belongs to the same cluster, which one of ordinary skill in the art would recognize could include more than one cluster. For example, one pair of items could belong to one cluster while another pair of items belongs to another cluster.)
the calculating includes calculating the Adjusted Rand Index as the predetermined evaluation scale based on the estimated contribution rate and the plurality of labels; and (Page 375 states "We test the performance of the discussed approaches to find an estimated clustering
c
^
that is close to the true clustering by fitting a simple Dirichlet process mixture model with normal components to simulated data and gene expression data." Page 373 states "The adjusted Rand index with the posterior expected clustering
A
R
c
*
,
E
c
y
is given by the expression [Expression (13)]." Page 374 states "Expression (13) requires the computation of the posterior similarity matrix, but can then be evaluated a lot faster than (12), which is advantageous if the criterion needs to be calculated for many different
c
*
. We also found (13) to be more amenable to a theoretical study. Expression (12) on the other hand does not require the computation of the posterior similarity matrix, which can be preferable for large n, where this matrix gets too large to be stored. And one can argue that maximizing
E
(
A
R
(
c
*
;
c
)
|
y
)
is a more standard approach than maximizing
A
R
c
*
,
E
c
y
. Practically, however, we found in our applications that maximization of either criterion leads to nearly identical results. For simplicity we will in the following refer to both criteria as PEAR for Posterior Expected Adjusted Rand." When fitting a model, one of ordinary skill in the art would recognize that multiple iterations of potential fittings are compared to one another using the loss function/criterion, which is in this case, the Adjusted Rand Index. Expression (13) shows the calculation of the loss function, which includes Expression (3), which calculates the contribution rate
P
c
i
=
c
j
y
)
. Additionally, as the labels represent the clusters that the items belong to, and the Adjusted Rand Index depends on the estimated and true clusters, the Adjusted Rand Index is based on the plurality of labels.)
maximizing an expected value of the Adjusted Rand Index (Page 373 states "The adjusted Rand index with the posterior expected clustering
A
R
c
*
,
E
c
y
is given by the expression [Expression (13)]." Page 374 states "Expression (13) requires the computation of the posterior similarity matrix, but can then be evaluated a lot faster than (12), which is advantageous if the criterion needs to be calculated for many different
c
*
. We also found (13) to be more amenable to a theoretical study. Expression (12) on the other hand does not require the computation of the posterior similarity matrix, which can be preferable for large n, where this matrix gets too large to be stored. And one can argue that maximizing
E
(
A
R
(
c
*
;
c
)
|
y
)
is a more standard approach than maximizing
A
R
c
*
,
E
c
y
. Practically, however, we found in our applications that maximization of either criterion leads to nearly identical results. For simplicity we will in the following refer to both criteria as PEAR for Posterior Expected Adjusted Rand." Maximizing PEAR/
A
R
c
*
,
E
c
y
/ Expression (13) is interpreted as the maximization of the expected value.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Allen and Fritsch for the reasons given above in regards to claim 1.
Claim(s) 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Allen (“Infinite Mixture Prototypes for Few-Shot Learning”, 2019) and Fritsch (“Improved Criteria for Clustering Based on the Posterior Similarity Matrix”, 2009) as applied to claim 1 above, further in view of Blei (“Variational inference for Dirichlet process mixtures”, 2004).
Regarding claim 10, the rejection of claim 1 is incorporated herein. Allen teaches
the learning is performed by learning a parameter of the neural network by maximizing an expected value of … using a gradient estimation method to optimize the neural network for generating the representation data that forms a cluster structure consistent with the plurality of labels. (Page 4 states "IMP optimizes the embedding parameters
ϕ
and cluster variances
σ
by stochastic gradient descent across episodes." As above, the models are optimized through the loss function, interpreted as the evaluation scale. As the neural network is able to produce the representation data by embedding the input data, the neural network is optimized for the embedding parameters which generate the representation. Page 2 states "For nonparametric representation learning methods, the model parameters are for the embedding function
h
ϕ
:
R
D
→
R
M
that map an input point
x
into a feature. The embedding of point
x
is the
M
-dimensional feature vector from the embedding function. In deep models the parameters
ϕ
are the weights of a deep network, and the embedding is the output of the last layer of this network." Page 10 states "For all few-shot experiments, we use the same base architecture as prototypical networks for the embedding network." Page 4, Algorithm 1 states “Init. each cluster
μ
c
with label
l
c
and
σ
c
=
σ
l
as class-wise means of the supports”. Page 3 states “Prototypical networks (Snell et al., 2017) form prototypes as the mean of the embedded support points in each class:
μ
n
=
1
|
S
n
|
∑
x
i
,
y
i
∈
S
n
h
ϕ
x
i
.” Thus, a neural network is used to convert the items of data into embeddings, interpreted as the representation data which forms a cluster structure consistent with the labels. )
Allen does not appear to explicitly teach
the clustering is performed by clustering the plurality of items of representation data by estimating a probability that each item of the representation data belongs to each cluster
the calculating is performed by calculating a non-differentiable clustering evaluation scale including an Adjusted Rand Index (ARI) based on the estimated probability and the plurality of labels; and
the non-differentiable clustering evaluation scale
However, Fritsch—directed to analogous art—teaches
the clustering is performed by clustering the plurality of items of representation data by estimating a probability that each item of the representation data belongs to each cluster (Page 369 states "If
K
varies between the clusterings a possible solution is to choose a clustering based on the posterior similarity matrix
P
c
i
=
c
j
y
)
, an
n
×
n
matrix that contains the pairwise probabilities that two observations belong to the same cluster. This approach is taken, for example, in the Bayesian cluster models for microarray data by Dahl (2006) and Medvedovic, Yeung, and Bumgarner (2004)."
P
c
i
=
c
j
y
)
is interpreted as the contribution rate, as it indicates the probabilities of each pair of items of representation data
i
,
j
, which includes each of the items of representation data belongs to the same cluster, which one of ordinary skill in the art would recognize could include more than one cluster. For example, one pair of items could belong to one cluster while another pair of items belongs to another cluster.)
the calculating is performed by calculating a non-differentiable clustering evaluation scale including an Adjusted Rand Index (ARI) based on the estimated probability and the plurality of labels; and (Page 375 states "We test the performance of the discussed approaches to find an estimated clustering
c
^
that is close to the true clustering by fitting a simple Dirichlet process mixture model with normal components to simulated data and gene expression data." Page 373 states "The adjusted Rand index with the posterior expected clustering
A
R
c
*
,
E
c
y
is given by the expression [Expression (13)]." Page 374 states "Expression (13) requires the computation of the posterior similarity matrix, but can then be evaluated a lot faster than (12), which is advantageous if the criterion needs to be calculated for many different
c
*
. We also found (13) to be more amenable to a theoretical study. Expression (12) on the other hand does not require the computation of the posterior similarity matrix, which can be preferable for large n, where this matrix gets too large to be stored. And one can argue that maximizing
E
(
A
R
(
c
*
;
c
)
|
y
)
is a more standard approach than maximizing
A
R
c
*
,
E
c
y
. Practically, however, we found in our applications that maximization of either criterion leads to nearly identical results. For simplicity we will in the following refer to both criteria as PEAR for Posterior Expected Adjusted Rand." When fitting a model, one of ordinary skill in the art would recognize that multiple iterations of potential fittings are compared to one another using the loss function/criterion, which is in this case, the Adjusted Rand Index, which is an inherently non-differentiable clustering evaluation scale. Expression (13) shows the calculation of the loss function, which includes Expression (3), which calculates the contribution rate
P
c
i
=
c
j
y
)
. Additionally, as the labels represent the clusters that the items belong to, and the Adjusted Rand Index depends on the estimated and true clusters, the Adjusted Rand Index is based on the plurality of labels.)
the non-differentiable clustering evaluation scale (As stated above, the Adjusted Rand Index is the non-differentiable clustering evaluation scale.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Allen and Fritsch for the reasons given above in regards to claim 1.
The combination of Allen and Fritsch does not appear to explicitly teach
using a variational Bayesian method
However, Blei—directed to analogous art—teaches
using a variational Bayesian method (Page 3 states "In this paper, we present a variational inference algorithm for DP mixtures based on the stick-breaking representation of the underlying DP."
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine the teachings of Allen and Fritsch with the teachings of Blei because, as Blei states on page 13, "Qualitatively, variational methods offer several potential advantages over Gibbs sampling. They are deterministic, and have an optimization criterion given by Equation (16) that can be used to assess convergence. In contrast, assessing convergence of a Gibbs sampler—namely, determining when the Markov chain has reached its stationary distribution—is an active field of research. Theoretical bounds on the mixing time are of little practical use, and there is no consensus on how to choose among the several empirical methods developed for this purpose (Robert and Casella 2004).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 JESSICA THUY PHAM whose telephone number is (571)272-2605. The examiner can normally be reached Monday - Friday, 9 A.M. - 5:00 P.M..
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Li Zhen can be reached at (571) 272-3768. 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.
/J.T.P./Examiner, Art Unit 2121 /Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights