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 Arguments
(Submitted 7/29/2026)
Regarding 101 Rejections
- The applicant on Page 8 traverses the rejections and argues on Page 9 that the claims do not recite a judicial exception and argues that claims 1, 19 and 20 are amended and that claim 1 is not directed to a mental step or data manipulation as the machine learning apparatus is extracting the feature. The applicant on Page 11 further argues that the claim may integrate a judicial exception into a practical application in a “meaningful way”.
Examiner’s Response:
The examiner first submits that there are NO claims 19 and 20 in this application as stated by the applicant. The examiner submits that the applicant has amended claims 1, 10 and 11. The examiner interprets that Variational Dropout is a stochastic regularization technique in neural networks where dropout masks are sampled randomly during training. This is a traditional computational procedure executed by a machine. The additional element does not amount to significantly more. (See 20200387782 which describes variational dropout known in the art, spec [0053])
In CONCLUSION, the examiner hereby MAINTAINS the 101 rejections on claims 1-2 and 4-12. The examiner notes that the applicant has CANCELLED claim 3.
Regarding 103 Rejections
- The applicant argues on Page 16 argues that the prior art does not teach the amended claims with combination including “Yaguchi” from claim 3 which is now “CACNCELED”.
Examiner’s Response:
Considering the broader context of amended limitations, the examiner submits that the applicant argument is MOOT and has used new grounds of rejections for claims 1, 10-11 and has used a new reference “Boluki” to teach the amendments. The examiner submits that a new rejection based on prior art that teaches that limitation and finds that the same new references also teach other limitations in the amended claim as relevant to the amended claim and meets the statutory grounds for rejection. The examiner also submits an examiner may cite a new reference in the independent claim to reject a dependent claim that has not been amended, provided the new reference is relevant to the dependent claim’s limitations and the rejection is supported by the claim language and the specification. The examiner submits that claim 6 is not amended but is taught by the new reference “Boluki” as it is directly relevant to the amended claim limitations.
In CONCLUSION, the examiner RETAINS the 103 rejections on claims 1-2 and 4-12 and MOVE the application to FINAL REJECTION under 103. The examiner notes that the applicant has CANCELLED claim 3.
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 1-2, and 4-12 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1:
According to the first part of the analysis, in the instant case, claim 1 is directed to a learning apparatus (machine) and claim 10 is directed to a learning method (method) and claim 11 is directed to a non-transitory media for leaning method (machine. Thus, each of the claims falls within one of the four statutory categories (i.e. process, machine, manufacture, or composition of matter).
Regarding claim 1:(Currently Amended)
Step 2A Prong 1:
“generate [[a]] first converted feature values and [[a]] second converted feature values” is a mental step of data manipulation.
Additional Elements
Step 2A Prong 2:
“ A machine learning apparatus comprising processing circuitry configured to: extract, using a machine learning model, first feature values from input data by processing based on one or more first parameters” recited in the preamble do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extract, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculate a first loss related to similarity between the first converted feature values and the second converted feature value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtain a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
“ and update a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” do not integrate the judicial exception into a practical application. These additional elements are merely directed to use a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ A machine learning apparatus comprising processing circuitry configured to: extract, using a machine learning model, first feature values from input data by processing based on one or more first parameters” recited in the preamble does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extract, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculate a first loss related to similarity between the first converted feature values and the second converted feature value” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtain a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and update a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 2: (Original)
Step 2A Prong 1:
“ by replacing at least one element of the first feature values and the second feature values selected by a random number with a predetermined value” is a mental step of data manipulation.
Additional Elements
Step 2A Prong 2:
“ wherein the processing circuitry generates the first converted feature values and the second converted feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processing circuitry generates the first converted feature values and the second converted feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 4: (Original)
Step 2A Prong 2:
“ wherein the processing circuitry generates the first converted feature values or the second converted feature values by a weighted average of the first feature values and the second feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processing circuitry generates the first converted feature values or the second converted feature values by a weighted average of the first feature values and the second feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 5: (Original)
Step 2A Prong 2:
“ wherein at least one of first processing and second processing is executed a plurality of times, the first processing performing processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the first feature values, the second processing performing processing of the processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the second feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein at least one of first processing and second processing is executed a plurality of times, the first processing performing processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the first feature values, the second processing performing processing of the processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the second feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 6: (Original)
Step 2A Prong 2:
“ wherein the processing circuitry executes preprocessing including one or more conversions on the input data” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processing circuitry executes preprocessing including one or more conversions on the input data” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 7: (Original)
Step 2A Prong 1:
“ and update a parameter of at least one of the second parameters and the fourth parameters such that a value based on a third loss calculated from the second processing result and a label is minimized” is a mental step of data manipulation.
Additional Elements
Step 2A Prong 2:
“wherein the processing circuitry is further configured to: obtain a second processing result by executing processing based on one or more fourth parameters different from the first to the third parameters, with respect to the second converted feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“wherein the processing circuitry is further configured to: obtain a second processing result by executing processing based on one or more fourth parameters different from the first to the third parameters, with respect to the second converted feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 8: (Original)
Step 2A Prong 2:
“ wherein the processing circuitry updates the first parameters and the third parameters after completion of updating of the second parameters and the fourth parameters” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processing circuitry updates the first parameters and the third parameters after completion of updating of the second parameters and the fourth parameters” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 9: (Original)
Step 2A Prong 2:
“ wherein the processing circuitry updates the second parameters based on updated first parameter” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processing circuitry updates the second parameters based on updated first parameter” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 10: (Currently Amended)
Step 2A Prong 1:
“generating [[a]] first converted feature values and [[a]] second converted feature values “is a mental step of data manipulation.
Additional Elements
Step 2A Prong 2:
“ A machine learning method, comprising : extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters” recited in the preamble do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculating a first loss related to similarity between the first converted feature values and the second converted feature value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
“ and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” do not integrate the judicial exception into a practical application. These additional elements are merely directed to use a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ A machine learning apparatus comprising processing circuitry configured to: extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters” recited in the preamble does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculating a first loss related to similarity between the first converted feature values and the second converted feature value” does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 11: (Currently Amended)
Step 2A Prong 1:
“generating [[a]] first converted feature values and [[a]] second converted feature values” is a mental step of data manipulation.
Additional Elements
Step 2A Prong 2:
“A non-transitory computer readable medium including computer executable Instructions, wherein the instructions, when executed by a processor, cause the processor to perform a machine method comprising:” recited in the preamble do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” do not integrate the judicial exception into a practical application. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculating a first loss related to similarity between the first converted feature values and the second converted feature value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
“and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” do not integrate the judicial exception into a practical application. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“A non-transitory computer readable medium including computer executable Instructions, wherein the instructions, when executed by a processor, cause the processor to perform a machine method comprising:” recited in the preamble do not integrate the judicial exception into a practical application” recited in the preamble does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“extracting, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ by using Variational DropOut or ZoneOut and by (a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values” does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“calculating a first loss related to similarity between the first converted feature values and the second converted feature value” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters”
does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized” does not amount to more than the judicial exception in the claim. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Regarding claim 12: (Original)
Step 2A Prong 1:
“the first network including a parameter group of the first parameters and the third parameters, the second network including a parameter group of the second parameters and the fourth parameters”(except the inference apparatus comprising processing circuitry configured to: “ is a mental step of topology construction using pen and paper.
Additional Elements
Step 2A Prong 2:
“ An inference apparatus using a first network and a second network trained by the learning according to claim 7” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ input processing target data to the first network and generate a first processing result; input processing target data to the second network and generate a second processing result” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and calculate at least one of a weighted average of the first processing result and the second processing result and reliability based on a difference between the first processing result and the second processing result” do not integrate the judicial exception into a practical application. These additional elements are merely directed at using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ An inference apparatus using a first network and a second network trained by the learning according to 8” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ input processing target data to the first network and generate a first processing result; input processing target data to the second network and generate a second processing result” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and calculate at least one of a weighted average of the first processing result and the second processing result and reliability based on a difference between the first processing result and the second processing result” does not amount to more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
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 1,6 and 9-11 are rejected under 35 U.S.C. 103 unpatentable over
Shahin Boluki et.al. (hereinafter Boluki) Learnable Bernoulli Dropout for Bayesian Deep Learning, Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108: 3905-3916, 2020,
[Note: As the article is not page numbered, the examiner has annotated with page number from the above citations]
in view of Hongsheng Wang et.al. (hereinafter Wang) US 11810366 B1.
Regarding claim 1: (Currently Amended)
Boluki discloses:
- A machine learning apparatus comprising processing circuitry configured to: extract, using a machine learning model, first feature values from input data by processing based on one or more first parameters
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ.
[2.1, Page 3906]:
dropout rates for input features for linear classifiers are tuned by a hierarchical Bayesian approach.
[BRI: in a neural network defined as f (x; θ), the parameters θ (weights and biases) determine how the network processes the input x to produce outputs, and this process can include learning feature representations from the raw input data
- generate a first converted feature values and a second converted feature values by stochastically converting at least one of the first feature values and the second feature values by using Variational DropOut
[Abstract, Page 3905]:
In this work, we propose learnable Bernoulli dropout (LBD), a new model-agnostic dropout scheme that considers the dropout rates as parameters jointly optimized with other model parameters. By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[2.1, Page 3907]:
We next formulate the problem of learning dropout rates for supervised feed-forward DNNs, and unsupervised latent representation learning in VAEs. In the formulations that follow, the dropout rates can be optimized using the method described in Section 2.4. We first briefly review the variational interpretation of dropout in Bayesian deep neural networks and show how our LBD fits here
[BRI: Variational dropout is a probabilistic extension of standard dropout that learns the dropout rate (or noise scale) instead of using a fixed hyperparameter, and it frames dropout within a variational Bayesian framework
- a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern,
[Abstract, Page 3905]:
By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[1, Page 3905]:
Considering the original dropout scheme as multiplying the output of each neuron by a binary mask drawn from a Bernoulli distribution, several dropout variations with other distributions for random multiplicative masks have been studied, including Gaussian dropout
[BRI: standard dropout is implemented by multiplying each neuron’s output by a binary mask drawn from a Bernoulli distribution, and there are many variants that use different distributions, patterns, or masking strategies. In standard neuron-level dropout, each neuron’s “drop” decision is modeled as an independent Bernoulli random variable, and in variational dropout this same Bernoulli mask is used as the stochastic “variational” factor that gates neuron contributions during training]
[3.2.2, Page 3910]:
Due to the concatenation of many features within each resolution, dropout becomes important to regularize the network by removing less relevant features from consideration. In our experiments, we replace the dropout in the 15 middle layers of this architecture with LBD as well as other forms of dropout. Here, the dropout parameters are shared between neurons in the same layer,
[2.3, Page 3908]:
variational distribution consists of an explicit conditional distribution q(η|
γ
) mixed with an implicit distribution q (
γ
). This allows the construction of a semi-implicit VAE (SIVAE), where the first stochastic layer,
q
ψ
(η|x), is still a Gaussian distribution, but other stochastic layers are constructed implicitly in the upper layers of the encoder.
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[3.3, Page 3913]:
the results for VAE without any dropout (VAE) are shown in Table 2. Note that because of the discrete (binary) nature of the implicit feedback data, Bernoulli dropout seems a naturally better fit for the input layers compared with other relaxed distributions due to not changing the nature of the data.
[BRI: in the standard Inverse Autoregressive Flow (IAF) formulation, each latent dimension is transformed using information from all previous dimensions, so the first and second feature values are stochastically coupled through the autoregressive dependency.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image3.png
18
415
media_image3.png
Greyscale
[BRI: Perhaps known to a POSITA that a sequence of binary random variables each parameterized by a Bernoulli probability represents a stochastic process, specifically a Bernoulli process, provided they are independent and identically distributed]
- calculate a first loss related to similarity between the first converted feature values and the second converted feature values.
[3.3, Page 3913]:
SLIM is a sparse linear model which learns an item-to-item similarity matrix by solving a constrained 1-regularization optimization.
[2.3, 3908]:
Regular dropout has been commonly used in training VAEs to prevent overfitting. When training, one sample is taken for each observed point in the mini-batch and the network parameters are found by minimizing
PNG
media_image4.png
133
485
media_image4.png
Greyscale
where the prior over the latent variables, p(η), is commonly set to N(0, I).
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[2.1, Page 3907]:
we propose to learn the dropout rates for Bernoulli masks jointly with the other model parameters. Specifically, we aim to optimize the expectation of the loss function with respect to the Bernoulli distribution dropouts:
PNG
media_image5.png
49
461
media_image5.png
Greyscale
[2.1, Page 3907]:
Under this formulation, we propose learnable Bernoulli dropout (LBD) as a variational approximation.
[2.4, Page 3908]:
In this paper, we estimate the gradient of (2) with respect to α with the ARM estimator
[2.4, Page 3908]:
Using ARM, we are able to directly optimize the Bernoulli dropout rates without introducing any bias.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image6.png
17
414
media_image6.png
Greyscale
With this gradient estimate, we can proceed to compute the gradient of our loss function in (1),
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ. The parameters are learned by minimizing an objective function L, which is usually comprised of an empirical loss E with possibly additional regularization terms R, by stochastic gradient descent (SGD):
PNG
media_image7.png
73
476
media_image7.png
Greyscale
where M is the mini-batch size.
[2.2, Page 3907]:
a prior is assumed over the weight matrices, p(W), which is updated to a posterior given the training data following Bayes’ rule p(W|D) = p(D|W)p(W)/ p(D) . This posterior distribution captures the set of plausible models and imposes a predictive distribution for the target of a new data point
[BRI: The “first converted feature values” is the observed binary user–item matrix and the “second converted feature values” is the predicted user–item matrix from the learned similarity matrix and the loss term is the direct measure of similarity quality in SLIM’s optimization. The rest of the objective (L1/L2 penalties) shapes sparsity and stability but does not replace this primary similarity loss]
Boluki does not explicitly disclose:
- extract, using a machine learning model, first feature values from the input data by processing based on one or more first parameters.
- extract, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters.
- obtain a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
- and update a parameter of at least one of the first parameters and the third parameters
such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
However, Wang discloses:
- extract, using a machine learning model, second feature from the input data by processing based on one or more second parameters different from the first parameters.
[Abstract]:
In the present disclosure, local features of pedestrians in video images are extracted by a multi-head attention neural network, weight parameters of image channels are learned by channel convolution kernels, spatial features on the images are scanned through spatial convolution, local features of pedestrians are enhanced to improve the recognition rate of pedestrians,
[Col 1, lines 49-51]:
The present disclosure provides a computer-implemented joint modeling method for enhancing local features of pedestrians
[Col 3, lines 32-37]:
an interactive concatenation of the multi-head attention neural network and the enhanced channel feature neural network: the output of the calculation for multi-head attention going through the first-layer three-channel convolution, learning different weight parameters,
[Col 10, lines 29-30]:
extract local image features of pedestrians; a module for extracting local features of pedestrians.
[Col 7, lines 63-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 1-16]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels
- obtain a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
[Col 3, lines 32-37]:
an interactive concatenation of the multi-head attention neural network and the enhanced channel feature neural network: the output of the calculation for multi-head attention going through the first-layer three-channel convolution, learning different weight parameters,
[Col 3, lines 65-67]; [Col 4, lines 1-16]:
adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S72: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
[BRI: the output is the result with the task (video surveillance application)]
[Col 7, lines 64-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 64-67]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels, and the calculation formula is as follows:
[Col 8, lines 1-42]:
PNG
media_image8.png
60
378
media_image8.png
Greyscale
where: O( i, j) is an output matrix, I is an input matrix, K is a convolution kernel matrix, and the convolution kernel matrix K has a shape of m x n; I (i+m, j+n) K (m,n) K(m,n) represents that elements of the input matrix I( i+m,j+n) are multiplied by elements of the kernel matrix K(m, n), and
PNG
media_image9.png
33
152
media_image9.png
Greyscale
is accumulated and summed in horizontal and vertical directions of the matrix, respectively; and Step 2: allowing the three convolution kernels to calculate independently, and to learn differential parameter weights among the three channels, and obtaining feature space maps of the three channels, which are then calculated interactively to obtain image channel features of pedestrians.
- and update a parameter of at least one of the first parameters and the third parameters
such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7.2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
[Col 10, lines 50-53]:
configured to construct a feed-forward neural network, where the enhanced local image features of pedestrians go through linear transformation and are then mapped into a pedestrian probability output; a module for model training, configured to iteratively train the neural network obtained by joint modeling and update model parameters until the model training converges,
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki, and Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Regarding claim 6: (Original)
Boluki discloses:
- wherein the processing circuitry executes preprocessing including one or more conversions on the input data.
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
[BRI: TensorFlow provides pre-processing by preparing the datasets for ML models]
Boluki and Wang do not explicitly disclose:
- wherein the processing circuitry executes preprocessing including one or more conversions on the input data.
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
[3.3, Page 3913]:
VAE-IAF uses invertible transformations based on an autoregressive neural network to build flexible variational posterior distributions.
[3.3, Page 3912]:
All the interactions of users in training set are used during training. For validation and test sets, 80% of interactions are randomly selected to learn the latent user-level representations, with the other 20% held-out to calculate the evaluation metrics for model predictions.
[BRI: TensorFlow provides preprocessing tools, primarily through TensorFlow Transform (TFX), which can normalize, encode, and transform input data for both training and serving and with TensorFlow and VAE-IAF the data can be compressed into a latent representation and can reconstruct the input]
Regarding claim 9: (Original)
Boluki does not explicitly disclose:
- wherein the processing circuitry updates the second parameters based on updated first parameters.
However, Wang discloses:
[Abstract]:
weight parameters of image channels are learned by channel convolution kernels, spatial features on the images are scanned through spatial convolution
[Col 3, lines 33-38]:
S62: an interactive concatenation of the multi-head attention neural network and the enhanced channel feature neural network: the output of the calculation for multi-head attention going through the first-layer three-channel convolution, learning different weight parameters among the three channels
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7. 2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample,
[Col 4, lines 63-67]:
a module for model training, configured to iteratively train the neural network obtained by joint modeling and update model parameters until the model training converges,
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki and Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Regarding claim 10: (Currently Amended)
Boluki discloses:
- A learning method, comprising extracting first feature values from input data by processing based on one or more first parameters.
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ.
[2.1, Page 3906]:
dropout rates for input features for linear classifiers are tuned by a hierarchical Bayesian approach.
[BRI: in a neural network defined as f (x; θ), the parameters θ (weights and biases) determine how the network processes the input x to produce outputs, and this process can include learning feature representations from the raw input data
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
- first feature values from input data by processing based on one or more first parameters
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ.
- generating [[a]] first converted feature values and a second converted feature values by stochastically converting at least one of the first feature values and the second feature values by using Variational DropOut
[Abstract, Page 3905]:
In this work, we propose learnable Bernoulli dropout (LBD), a new model-agnostic dropout scheme that considers the dropout rates as parameters jointly optimized with other model parameters. By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[2.1, Page 3907]:
We next formulate the problem of learning dropout rates for supervised feed-forward DNNs, and unsupervised latent representation learning in VAEs. In the formulations that follow, the dropout rates can be optimized using the method described in Section 2.4. We first briefly review the variational interpretation of dropout in Bayesian deep neural networks and show how our LBD fits here
[BRI: Variational dropout is a probabilistic extension of standard dropout that learns the dropout rate (or noise scale) instead of using a fixed hyperparameter, and it frames dropout within a variational Bayesian framework
- a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern,
[Abstract, Page 3905]:
By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[1, Page 3905]:
Considering the original dropout scheme as multiplying the output of each neuron by a binary mask drawn from a Bernoulli distribution, several dropout variations with other distributions for random multiplicative masks have been studied, including Gaussian dropout
[BRI: standard dropout is implemented by multiplying each neuron’s output by a binary mask drawn from a Bernoulli distribution, and there are many variants that use different distributions, patterns, or masking strategies. In standard neuron-level dropout, each neuron’s “drop” decision is modeled as an independent Bernoulli random variable, and in variational dropout this same Bernoulli mask is used as the stochastic “variational” factor that gates neuron contributions during training]
[3.2.2, Page 3910]:
Due to the concatenation of many features within each resolution, dropout becomes important to regularize the network by removing less relevant features from consideration. In our experiments, we replace the dropout in the 15 middle layers of this architecture with LBD as well as other forms of dropout. Here, the dropout parameters are shared between neurons in the same layer,
[2.3, Page 3908]:
variational distribution consists of an explicit conditional distribution q(η|
γ
) mixed with an implicit distribution q (
γ
). This allows the construction of a semi-implicit VAE (SIVAE), where the first stochastic layer,
q
ψ
(η|x), is still a Gaussian distribution, but other stochastic layers are constructed implicitly in the upper layers of the encoder.
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[3.3, Page 3913]:
the results for VAE without any dropout (VAE) are shown in Table 2. Note that because of the discrete (binary) nature of the implicit feedback data, Bernoulli dropout seems a naturally better fit for the input layers compared with other relaxed distributions due to not changing the nature of the data.
[BRI: in the standard Inverse Autoregressive Flow (IAF) formulation, each latent dimension is transformed using information from all previous dimensions, so the first and second feature values are stochastically coupled through the autoregressive dependency.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image3.png
18
415
media_image3.png
Greyscale
[BRI: Perhaps known to POSITA that a sequence of binary random variables each parameterized by a Bernoulli probability represents a stochastic process, specifically a Bernoulli process, provided they are independent and identically distributed]
- calculate a first loss related to similarity between the first converted feature values and the second converted feature values.
[3.3, Page 3913]:
SLIM is a sparse linear model which learns an item-to-item similarity matrix by solving a constrained 1-regularization optimization.
[2.3, 3908]:
Regular dropout has been commonly used in training VAEs to prevent overfitting. When training, one sample is taken for each observed point in the mini-batch and the network parameters are found by minimizing
PNG
media_image4.png
133
485
media_image4.png
Greyscale
where the prior over the latent variables, p(η), is commonly set to N(0, I).
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[2.1, Page 3907]:
we propose to learn the dropout rates for Bernoulli masks jointly with the other model parameters. Specifically, we aim to optimize the expectation of the loss function with respect to the Bernoulli distribution dropouts:
PNG
media_image5.png
49
461
media_image5.png
Greyscale
[2.1, Page 3907]:
Under this formulation, we propose learnable Bernoulli dropout (LBD) as a variational approximation.
[2.4, Page 3908]:
In this paper, we estimate the gradient of (2) with respect to α with the ARM estimator
[2.4, Page 3908]:
Using ARM, we are able to directly optimize the Bernoulli dropout rates without introducing any bias.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image6.png
17
414
media_image6.png
Greyscale
With this gradient estimate, we can proceed to compute the gradient of our loss function in (1),
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ. The parameters are learned by minimizing an objective function L, which is usually comprised of an empirical loss E with possibly additional regularization terms R, by stochastic gradient descent (SGD):
PNG
media_image7.png
73
476
media_image7.png
Greyscale
where M is the mini-batch size.
[2.2, Page 3907]:
a prior is assumed over the weight matrices, p(W), which is updated to a posterior given the training data following Bayes’ rule p(W|D) = p(D|W)p(W)/ p(D) . This posterior distribution captures the set of plausible models and imposes a predictive distribution for the target of a new data point
[BRI: The “first converted feature values” is the observed binary user–item matrix and the “second converted feature values” is the predicted user–item matrix from the learned similarity matrix and the loss term is the direct measure of similarity quality in SLIM’s optimization. The rest of the objective (L1/L2 penalties) shapes sparsity and stability but does not replace this primary similarity loss]
Boluki does not explicitly disclose:
- extracting, using a machine learning model, first feature values from the input data by processing based on one or more first parameters.
- extracting, using a machine learning model, second feature values from the input data by processing based on one or more second parameters different from the first parameters.
- obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
- and updating a parameter of at least one of the first parameters and the third parameters
such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
However, Wang discloses:
- extracting, using a machine learning model, first feature values from the input data by processing based on one or more first parameters.
[Col 7, lines 64-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 64-67]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels, and the calculation formula is as follows:
[Col 8, lines 1-42]:
PNG
media_image8.png
60
378
media_image8.png
Greyscale
where: O( i, j) is an output matrix, I is an input matrix, K is a convolution kernel matrix, and the convolution kernel matrix K has a shape of m x n; I (i+m, j+n) K (m,n) K(m,n) represents that elements of the input matrix I( i+m, j+n) are multiplied by elements of the kernel matrix K(m, n), and
PNG
media_image9.png
33
152
media_image9.png
Greyscale
is accumulated and summed in horizontal and vertical directions of the matrix, respectively; and Step 2: allowing the three convolution kernels to calculate independently, and to learn differential parameter weights among the three channels, and obtaining feature space maps of the three channels, which are then calculated interactively to obtain image channel features of pedestrians.
- extracting, using a machine learning model, second feature from the input data by processing based on one or more second parameters different from the first parameters.
[Abstract]:
In the present disclosure, local features of pedestrians in video images are extracted by a multi-head attention neural network, weight parameters of image channels are learned by channel convolution kernels, spatial features on the images are scanned through spatial convolution, local features of pedestrians are enhanced to improve the recognition rate of pedestrians,
[Col 1, lines 49-51]:
The present disclosure provides a computer-implemented joint modeling method for enhancing local features of pedestrians
[Col 3, lines 32-37]:
an interactive concatenation of the multi-head attention neural network and the enhanced channel feature neural network: the output of the calculation for multi-head attention going through the first-layer three-channel convolution, learning different weight parameters,
[Col 10, lines 29-30]:
extract local image features of pedestrians; a module for extracting local features of pedestrians.
[Col 7, lines 63-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 1-16]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels
- obtaining a first processing result of a task of the machine learning model by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
[Col 3, lines 32-37]:
an interactive concatenation of the multi-head attention neural network and the enhanced channel feature neural network: the output of the calculation for multi-head attention going through the first-layer three-channel convolution, learning different weight parameters,
[Col 3, lines 65-67]; [Col 4, lines 1-16]:
adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S72: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
[BRI: the output is the result with the task (video surveillance application)]
[Col 7, lines 64-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 64-67]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels, and the calculation formula is as follows:
[Col 8, lines 1-42]:
PNG
media_image8.png
60
378
media_image8.png
Greyscale
where: O( i, j) is an output matrix, I is an input matrix, K is a convolution kernel matrix, and the convolution kernel matrix K has a shape of m x n; I (i+m, j+n) K (m,n) K(m,n) represents that elements of the input matrix I( i+m,j+n) are multiplied by elements of the kernel matrix K(m, n), and
PNG
media_image9.png
33
152
media_image9.png
Greyscale
is accumulated and summed in horizontal and vertical directions of the matrix, respectively; and Step 2: allowing the three convolution kernels to calculate independently, and to learn differential parameter weights among the three channels, and obtaining feature space maps of the three channels, which are then calculated interactively to obtain image channel features of pedestrians.
- and updating a parameter of at least one of the first parameters and the third parameters
such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7.2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
[Col 10, lines 50-53]:
configured to construct a feed-forward neural network, where the enhanced local image features of pedestrians go through linear transformation and are then mapped into a pedestrian probability output; a module for model training, configured to iteratively train the neural network obtained by joint modeling and update model parameters until the model training converges,
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki, and Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Regarding claim 11(Currently Amended)
Boluki discloses:
- extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters;
[3, Page 3909]:
we investigate the performance on a toy example, where we can calculate the true gradients exactly, to demonstrate the advantages of LBD over the existing dropout schema. We implement all of our methods in Tensorflow (Abadi et al., 2016) on a single cluster node with Intel Xeon E5-2680 v4 2.40 GHz processor
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ.
[2.1, Page 3906]:
dropout rates for input features for linear classifiers are tuned by a hierarchical Bayesian approach.
[BRI: in a neural network defined as f (x; θ), the parameters θ (weights and biases) determine how the network processes the input x to produce outputs, and this process can include learning feature representations from the raw input data
- generating [a]] first converted feature values and a second converted feature values by stochastically converting at least one of the first feature values and the second feature values
[Abstract, Page 3905]:
In this work, we propose learnable Bernoulli dropout (LBD), a new model-agnostic dropout scheme that considers the dropout rates as parameters jointly optimized with other model parameters. By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[2.1, Page 3907]:
We next formulate the problem of learning dropout rates for supervised feed-forward DNNs, and unsupervised latent representation learning in VAEs. In the formulations that follow, the dropout rates can be optimized using the method described in Section 2.4. We first briefly review the variational interpretation of dropout in Bayesian deep neural networks and show how our LBD fits here
[BRI: Variational dropout is a probabilistic extension of standard dropout that learns the dropout rate (or noise scale) instead of using a fixed hyperparameter, and it frames dropout within a variational Bayesian framework
- a) adding at least one of the first feature values and the second feature values to a pattern generated based on a random number or (b) multiplying at least one of the first feature values and the second feature values by the pattern, to stochastically convert at least one of the first feature values and the second feature values
[Abstract, Page 3905]:
By probabilistic modeling of Bernoulli dropout, our method enables more robust prediction and uncertainty quantification in deep models. Especially, when combined with variational auto-encoders (VAEs), LBD enables flexible semi-implicit posterior representations, leading to new semi-implicit VAE (SIVAE) models.
[1, Page 3905]:
Considering the original dropout scheme as multiplying the output of each neuron by a binary mask drawn from a Bernoulli distribution, several dropout variations with other distributions for random multiplicative masks have been studied, including Gaussian dropout
[BRI: standard dropout is implemented by multiplying each neuron’s output by a binary mask drawn from a Bernoulli distribution, and there are many variants that use different distributions, patterns, or masking strategies. In standard neuron-level dropout, each neuron’s “drop” decision is modeled as an independent Bernoulli random variable, and in variational dropout this same Bernoulli mask is used as the stochastic “variational” factor that gates neuron contributions during training]
[3.2.2, Page 3910]:
Due to the concatenation of many features within each resolution, dropout becomes important to regularize the network by removing less relevant features from consideration. In our experiments, we replace the dropout in the 15 middle layers of this architecture with LBD as well as other forms of dropout. Here, the dropout parameters are shared between neurons in the same layer,
[2.3, Page 3908]:
variational distribution consists of an explicit conditional distribution q(η|
γ
) mixed with an implicit distribution q (
γ
). This allows the construction of a semi-implicit VAE (SIVAE), where the first stochastic layer,
q
ψ
(η|x), is still a Gaussian distribution, but other stochastic layers are constructed implicitly in the upper layers of the encoder.
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[3.3, Page 3913]:
the results for VAE without any dropout (VAE) are shown in Table 2. Note that because of the discrete (binary) nature of the implicit feedback data, Bernoulli dropout seems a naturally better fit for the input layers compared with other relaxed distributions due to not changing the nature of the data.
[BRI: in the standard Inverse Autoregressive Flow (IAF) formulation, each latent dimension is transformed using information from all previous dimensions, so the first and second feature values are stochastically coupled through the autoregressive dependency.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image3.png
18
415
media_image3.png
Greyscale
[BRI: Perhaps known to POSITA that a sequence of binary random variables each parameterized by a Bernoulli probability represents a stochastic process, specifically a Bernoulli process, provided they are independent and identically distributed]
- calculating a first loss related to similarity between the first converted feature values and the second converted feature values.
[3.3, Page 3913]:
SLIM is a sparse linear model which learns an item-to-item similarity matrix by solving a constrained 1-regularization optimization.
[2.3, 3908]:
Regular dropout has been commonly used in training VAEs to prevent overfitting. When training, one sample is taken for each observed point in the mini-batch and the network parameters are found by minimizing
PNG
media_image4.png
133
485
media_image4.png
Greyscale
where the prior over the latent variables, p(η), is commonly set to N(0, I).
[2.3, Page 3908]:
the stochastic layers were constructed by concatenating the deterministic layer output with random noise and observed data to form the input to the next layer.
[2.1, Page 3907]:
we propose to learn the dropout rates for Bernoulli masks jointly with the other model parameters. Specifically, we aim to optimize the expectation of the loss function with respect to the Bernoulli distribution dropouts:
PNG
media_image5.png
49
461
media_image5.png
Greyscale
[2.1, Page 3907]:
Under this formulation, we propose learnable Bernoulli dropout (LBD) as a variational approximation.
[2.4, Page 3908]:
In this paper, we estimate the gradient of (2) with respect to α with the ARM estimator
[2.4, Page 3908]:
Using ARM, we are able to directly optimize the Bernoulli dropout rates without introducing any bias.
[2.4, Page 3908]:
For a vector of K binary random variables z = [
Z
1
,
Z
2
,
…
Z
K
] parameterized by α = [
[
α
1
,
α
2
,
…
α
K
], the logits of the Bernoulli probability parameters, the gradient of
E
z
[L(θ, z)] with respect to α can be expressed as (Yin and Zhou, 2019)
PNG
media_image1.png
117
493
media_image1.png
Greyscale
Here Unif denotes the uniform distribution, and
PNG
media_image2.png
28
124
media_image2.png
Greyscale
denotes the loss obtained by setting
PNG
media_image6.png
17
414
media_image6.png
Greyscale
With this gradient estimate, we can proceed to compute the gradient of our loss function in (1),
[2.1, Page 3906]: Learnable Bernoulli Dropout (LBD)
Given a training dataset D =
x
i
,
y
i
)
}
i
=
1
N
, where x and y denote the input and target of interest respectively, a neural network is a function f(x; θ) from the input space to the target space with parameters θ. The parameters are learned by minimizing an objective function L, which is usually comprised of an empirical loss E with possibly additional regularization terms R, by stochastic gradient descent (SGD):
PNG
media_image7.png
73
476
media_image7.png
Greyscale
where M is the mini-batch size.
[2.2, Page 3907]:
a prior is assumed over the weight matrices, p(W), which is updated to a posterior given the training data following Bayes’ rule p(W|D) = p(D|W)p(W)/ p(D) . This posterior distribution captures the set of plausible models and imposes a predictive distribution for the target of a new data point
[BRI: The “first converted feature values” is the observed binary user–item matrix and the “second converted feature values” is the predicted user–item matrix from the learned similarity matrix and the loss term is the direct measure of similarity quality in SLIM’s optimization. The rest of the objective (L1/L2 penalties) shapes sparsity and stability but does not replace this primary similarity loss]
Boluki does not explicitly disclose:
- A non-transitory computer readable medium including computer executable instructions, wherein the instructions, when executed by a processor, cause the processor to perform a method comprising:
- extracting, using a machine learning model, second feature values from input data by processing based on one or more second parameters different from the first parameters.
- obtaining a first processing result by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
- and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
However, Wang discloses:
- A non-transitory computer readable medium including computer executable instructions, wherein the instructions, when executed by a processor, cause the processor to perform a method comprising:
[Col 11, lines 58-67], in [Col 12, lines 1-5]:
- extracting, using a machine learning model, first feature values from input data by processing based on one or more first parameters.
[Col 10, lines 29-30]:
extract local image features of pedestrians; a module for extracting local features of pedestrians.
[Col 7, lines 63-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
In [Col 8, lines 1-16]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels
- extracting, using a machine learning model, second feature values from input data by processing based on one or more second parameters different from the first parameters
[Col 7, lines 64-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 64-67]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels, and the calculation formula is as follows:
[Col 8, lines 1-42]:
PNG
media_image8.png
60
378
media_image8.png
Greyscale
where: O( i, j) is an output matrix, I is an input matrix, K is a convolution kernel matrix, and the convolution kernel matrix K has a shape of m x n; I (i+m, j+n) K (m,n) K(m,n) represents that elements of the input matrix I( i+m, j+n) are multiplied by elements of the kernel matrix K(m, n), and
PNG
media_image9.png
33
152
media_image9.png
Greyscale
is accumulated and summed in horizontal and vertical directions of the matrix, respectively; and Step 2: allowing the three convolution kernels to calculate independently, and to learn differential parameter weights among the three channels, and obtaining feature space maps of the three channels, which are then calculated interactively to obtain image channel features of pedestrians.
- obtaining a first processing result by processing based on one or more third parameters with respect to the first converted feature values, the one or more third parameters being different from the first parameters and the second parameters.
[Col 7, lines 64-67]:
inputting the images into the three-channel image convolutional neural network to capture image channel features of pedestrians, and referring to FIG. 3, the fourth stage is divided into the following sub-steps:
[Col 8, lines 64-67]:
Step 1: with regard to the three channels of the input images, constructing a three-channel image convolutional neural network which includes three convolution kernels that correspond to the three channels of the images, respectively; allowing the three convolution kernels to learn weight parameters of corresponding image channels respectively, so as to output three groups of different weight parameters, each of the convolution kernels having a size of 1×1×3, where 3 is the number of channels of the input images; inputting the images into the three-channel image convolutional neural network, where the input images are weighted and combined in a convolution depth direction and, after going through the three convolution kernels of 1×1×3, output three local features which contain the weight parameters among the three channels, and the calculation formula is as follows:
[Col 8, lines 1-42]:
PNG
media_image8.png
60
378
media_image8.png
Greyscale
where: O( i, j) is an output matrix, I is an input matrix, K is a convolution kernel matrix, and the convolution kernel matrix K has a shape of m x n; I (i+m, j+n) K (m,n) K(m,n) represents that elements of the input matrix I( i+m, j+n) are multiplied by elements of the kernel matrix K(m, n), and
PNG
media_image9.png
33
152
media_image9.png
Greyscale
is accumulated and summed in horizontal and vertical directions of the matrix, respectively; and Step 2: allowing the three convolution kernels to calculate independently, and to learn differential parameter weights among the three channels, and obtaining feature space maps of the three channels, which are then calculated interactively to obtain image channel features of pedestrians.
- and updating a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7.2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
[Col 10, lines 50-53]:
configured to construct a feed-forward neural network, where the enhanced local image features of pedestrians go through linear transformation and are then mapped into a pedestrian probability output; a module for model training, configured to iteratively train the neural network obtained by joint modeling and update model parameters until the model training converges,
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki, and Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Claim 2 is rejected under 35 U.S.C. 103 unpatentable over
Shahin Boluki et.al. (hereinafter Boluki) Learnable Bernoulli Dropout for Bayesian Deep Learning, Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108: 3905-3916, 2020.
in view of Hongsheng Wang et.al. (hereinafter Wang) US 11810366 B1.
further in view of Utkarsh Ojha et.al. (hereinafter Ojha) US 2022/0254071 A1,
further in view of ROMA Cambray et.al. (hereinafter Cambray) US 2023/0197248 A1.
Regarding claim 2: (Original)
Boluki and Wang do not explicitly disclose:
- wherein the processing circuitry generates the first converted feature values and the second converted feature values by replacing at least one element of the first feature values and the second feature values selected by a random number with a predetermined value
However, Ojha discloses:
- wherein the processing circuitry generates the first converted feature values and the second converted feature values by replacing at least one element of the first feature values and the second feature values selected by a random number with a predetermined value
[0075]:
To define the anchor region 406, the GAN translation system 102 selects k random points (e.g., corresponding to the number of example digital images within a few-shot target set 408) within the latent space 402. In addition, the GAN translation system 102 samples from these fixed points with a small added Gaussian noise (σ=0.5). Further, the GAN translation system 102 utilizes shared parameters (e.g., weights) between the two discriminators by defining D.sub.patch as a subset of the larger D.sub.img neural network. In various embodiments, network size depends on the network architecture and layer. In some cases, the GAN translation system 102 reads off a set of layers with effective patch size ranging from 22×22 to 61×61.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, Wang and Ojha.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
Ojha teaches selecting a random number with a predetermined value
One of ordinary skills would am motivated to combine Boluki, Wang and Ojha that can improve the accuracy of the GAN using GAN translation (Ojha [0086]).
Boluki, Wang and Ojha do not explicitly disclose:
- wherein the processing circuitry generates the first converted feature values and the second converted feature values by replacing at least one element of the first feature values and the second feature values selected by a random number with a predetermined value
However, Cambray discloses:
- wherein the processing circuitry generates the first converted feature values and the second converted feature values by replacing at least one element of the first feature values and
[0081]:
In an embodiment of this invention, the CADx device (120) incorporates the ability to adapt to the medical image(s) scan protocol information (116). In some cases, it is expected that part or all of the scan protocol information (116) may be missing. It would therefore be advantageous to incorporate the ability to handle such scenario, where the scan protocol information is incomplete.
[0082]:
a specific point in that encoding space such as for example, the zero vector, may be chosen to represent a missing scan protocol parameter. During training, the stochastic drop-out step (801) sets protocol parameters as missing according to some policy, preferably by a random process with a pre-determined probability for each scan protocol parameter. When a scan protocol parameter is removed, it is simply substituted by the zero-vector. Note that outside of training, at the time of making predictions, this stochastic drop-out step is not applied and instead, the model is able to handle missing parameters by converting the missing parameter(s) to the zero-vector.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, Wang, and Cambray.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches first loss.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
Cambray teaches replacing an element of the feature values selected by a random number with a predetermined value
One of ordinary skill would be motivated to combine Boluki, Wang, Ojha and Cambray
that can provide minimum error metric using trainable machine learning model (Cambray [ 0059]).
Claims 4-5, 7-8 and 12 are rejected under 35 U.S.C. 103 unpatentable over
Shahin Boluki et.al. (hereinafter Boluki) Learnable Bernoulli Dropout for Bayesian Deep Learning, Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108: 3905-3916, 2020,
[Note: As the article is not page numbered, the examiner has annotated with page number from the above citations]
in view of Hongsheng Wang et.al. (hereinafter Wang) US 11810366 B1.
further Masahiro OGINO et.al. (hereinafter OGINO) US 2021/0272277A1.
Regarding claim 4: (Original)
Boluki, and Wang do not explicitly disclose:
- wherein the processing circuitry generates the first converted feature values or the second converted feature values by a weighted average of the first feature values and the second feature values.
However, OGINO discloses:
- wherein the processing circuitry generates the first converted feature values or the second converted feature values by a weighted average of the first feature values and the second feature values.
[0095]:
in FIG. 15 illustrating a feature quantity space, an error function in which weights are given to a distance between medically important teacher data A1 and an output B1 and a distance between next important teacher data group (data set S) Ai and an output Bi is set. When a conversion function to the feature quantity space ε is set to g and the center of gravity (average value of coordinates of each piece of data) on the space ε is represented by C, the error function is represented by the following Formula (2).
PNG
media_image10.png
92
375
media_image10.png
Greyscale
[0093]:
This error function is a combination of the above-mentioned predetermined spatial distance error and medical knowledge. The predetermined spatial distance error defines an error function that brings the entire space closer, using a center of gravity of a feature quantity space as a parameter. In this error function, a space to be matched is weighted based on medical knowledge and importance
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, Wang and OGINO.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
OGINO teaches weighted average of the feature values.
One of ordinary skill would be motivated to combine Boluki, Wang and OGINO that can provide improved accuracy of the predictive model (OGINO [0077])
Regarding claim 5: (Original)
Boluki, and Wang do not explicitly disclose:
- wherein at least one of first processing and second processing is executed a plurality of times, the first processing performing processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the first feature values, the second processing performing processing of the processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the second feature values.
However, OGINO discloses:
- wherein at least one of first processing and second processing is executed a plurality of times, the first processing performing processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the first feature values, the second processing performing processing of the processing of generation of the first converted feature values and the second converted feature values after processing of extraction of the second feature values.
[Abstract]:
The image processing unit includes a feature quantity extraction unit that extracts a first feature quantity from the first image data, a feature quantity abstraction unit that abstracts the first feature quantity to extract a second feature quantity, a feature quantity conversion unit that converts the second feature quantity into a third feature quantity extracted by second image data, and an identification unit that uses the converted third feature quantity to calculate a predetermined parameter value.
[0028]:
FIG. 14 is a diagram for description of a relationship between feature quantities in the feature quantity space;
PNG
media_image11.png
580
785
media_image11.png
Greyscale
[0058]:
The feature quantity A, which is the output of the predictive model 232M, expresses a plurality of classifications necessary for diagnosis of a feature of an image as a vector of a plurality of dimensions (for example, 1,024 dimensions), and a feature related to a parameter (for example, whether a tumor is benign or malignant) is extracted. Such a feature quantity A is obtained for each patch. Note that in FIG. 4, the feature quantity A 410 is an output of a final layer of all the combined layers. However, the invention is not limited thereto. Even though the deeper the layer, the greater the degree of feature abstraction, it is possible to use an output of a layer shallower than the final layer as the feature quantity.
[0055]:
As schematically illustrated in FIG. 4, a CNN 40 of the predictive model 232M is a computing unit constructed on a computer configured to repeat a convolution operation and pooling 43 a plurality of times between an input layer 41 and an output layer 44 on a multi-layer network. In FIG. 4, a number in front of a block indicating each layer is the number of layers, and a number in each layer represents a size processed by each layer. The CNN of this predictive model is learned to extract a feature quantity A 410 for accurately identifying the presence or absence of the lesion of an input image 400 by the CNN repeating the convolution calculation and pooling on input data of the input image 400 for learning divided into a plurality of patches by the patch processing unit 231.
[0061]:
The predictive model 233M receives a feature quantity corresponding to the number of patches output from the feature quantity extraction unit 232 as an input, and extracts a main feature quantity
[0056]:
Learning is performed until an error between an output and teacher data falls within a predetermined range. An error function used at this time will be described after the In [0053]:
the feature quantity extraction unit 232 extracts the feature quantity A from image data of an input image, a second model is a model for the feature quantity abstraction unit 233 to extract the feature quantity B abstracted from the feature quantity A, a third model is a feature quantity conversion model for the feature quantity conversion unit 234 to convert the feature quantity B into the feature quantity C,
(BRI: the conversion to C after extracting A and B)
[0120]:
as illustrated in FIG. 20, for example, a relevant part of the lesion is cut out by a patch from a plurality of T1 weighted images of the MRI to extract the feature quantity A, and each feature quantity A obtained by each patch is combined to extract the feature quantity B
[0108]:
the input image may be generated from the signal collected by the imaging unit 100, and the feature quantity A and the feature quantity B extracted from the input image can be converted into the feature quantity C of the image having more detailed information to calculate the parameter value used for more accurate diagnosis from the feature quantity C. In this way, it is possible to present more accurate diagnosis support information using the medical imaging apparatus.
[0074]:
The identification model 235M is incorporated in the identification unit 235 such that such a CNN is trained using a plurality of combinations of the feature quantity (feature quantity C) after conversion
[0129]:
Further, the conversion of the feature quantity in the feature quantity conversion unit 234 is not limited to two captured images, and can be applied to a plurality of different types of captured images.
[0095]:
For example, in FIG. 15 illustrating a feature quantity space, an error function in which weights are given to a distance between medically important teacher data A1 and an output B1 and a distance between next important teacher data group (data set S) Ai and an output Bi is set. When a conversion function to the feature quantity space ε is set to g and the center of gravity (average value of coordinates of each piece of data) on the space ε is represented by C, the error function is represented by the following Formula (6)
PNG
media_image12.png
107
363
media_image12.png
Greyscale
[0096] Here, α, β, and γ are weighting factors, for example, α=0.5, β=0.4, and γ=0.1.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, Wang and OGINO.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
OGINO teaches weighted average of the feature values.
One of ordinary skill would be motivated to combine Boluki, Wang and OGINO that can provide improved accuracy of the predictive model (OGINO [0077])
Regarding claim 7: (Original)
Boluki does not explicitly disclose:
- and update a parameter of at least one of the second parameters and the fourth parameters such that a value based on a third loss calculated from the second processing result and a label is minimized.
However, Wang discloses:
- and update a parameter of at least one of the second parameters and the fourth parameters such that a value based on a third loss calculated from the second processing result and a label is minimized.
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7. 2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki, Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Boluki, and Wang do not explicitly disclose:
- wherein the processing circuitry is further configured to: obtain a second processing result by executing processing based on one or more fourth parameters different from the first to the third parameters, with respect to the second converted feature values.
However, OGINO discloses:
- wherein the processing circuitry is further configured to: obtain a second processing result by executing processing based on one or more fourth parameters different from the first to the third parameters, with respect to the second converted feature values.
[0013]:
The image processing unit includes a feature quantity extraction unit that extracts a first feature quantity from the first image data, a feature quantity abstraction unit that extracts (abstracts) a more important second feature quantity from the first feature quantity, a feature quantity conversion unit that converts the second feature quantity into a third feature quantity extracted by second image data different from the first image data, and an identification unit that uses the converted third feature quantity to calculate a predetermined parameter value, and performs prediction.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, Wang and OGINO.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
OGINO teaches weighted average of the feature values.
One of ordinary skill would be motivated to combine Boluki, Wang and OGINO that can provide improved accuracy of the predictive model (OGINO [0077])
Regarding claim 8: (Original)
Boluki does not explicitly disclose:
- wherein the processing circuitry updates the first parameters and the third parameters after completion of updating of the second parameters and the fourth parameters.
However, Wang discloses:
- wherein the processing circuitry updates the first parameters and the third parameters after completion of updating of the second parameters and the fourth parameters.
[Col 14, lines 37-54]:
S7.1: adopting the feed-forward neural network and the activation function, inputting the obtained enhanced local image features of pedestrians into the feed-forward neural network, allowing it to go through linear layer transformation, and mapping the probability distribution of pedestrians into categories to recognize pedestrians by using the activation function; and S7. 2: calculating an intersection ratio of coordinates of the recognized pedestrians and the image labeled sample in the original surveillance video image data set, and calculating an accuracy rate and a recall rate, where the accuracy rate refers to the recognized pedestrians, indicating a proportion of real pedestrians in the sample predicted to be positive, and the recall rate refers to the image labeled sample in the original surveillance video image data set, indicating a proportion of correctly recognized pedestrians in the positive examples in the sample.
It would have been obvious to one of ordinary skills in the art before the effective filing date of the present application to combine Boluki, and Wang.
Boluki teaches extracting features and obtaining converted feature values by stochastically converting the feature values by using variational drop out and teaches loss related to similarity of converted features.
Wang teaches processing the feature parameters and updating parameters to minimize the loss.
One of ordinary skill would be motivated to combine Boluki, and Wang that can enhance local features to improve the recognition rate of the object (pedestrians) (Wang [Col 5, lines 18-19])
Regarding claim 12: (Original)
Boluki and Wang do not explicitly disclose:
- An inference apparatus using a first network and a second network trained by the learning apparatus
However, OGINO discloses:
- An inference apparatus using a first network and a second network trained by the learning apparatus
[Abstract]:
A medical imaging apparatus includes an imaging unit that collects an image signal of an inspection target, and an image processing unit that generates first image data from the image signal and performs image processing of the first image data.
[Abstract]:
To obtain a predictive model that shows a diagnostic prediction result with higher accuracy and high medical validity.
[0063]:
The CNN is trained so that a feature quantity that most contributes to the parameter is output, and is used as the predictive model 233M of the feature quantity abstraction unit 233.
- the first network including a parameter group of the first parameters and the third parameters
[0118]:
a plurality of patterns of learning models such as a learning model according to an examination site of the subject
[0093]:
in the feature quantity space (feature quantity map), a relationship in image data among teacher data 1002, image data before conversion 1001 and after conversion 1003 from the feature quantity B to the feature quantity C is analyzed. For example, a feature quantity space error reduction is particularly weighted to the distance between a patch image (group) 1402 having a deep relation to determination of the presence or absence of a disease in a pathological image and an MRI image (group) 1401 corresponding thereto.
[0104]:
the feature quantity conversion unit 234 uses the feature quantity conversion model 234M (FIG. 6) to convert the feature quantity B into the feature quantity C (S5). The identification unit 235 uses the identification model 235M to calculate a parameter value for predicting
- the second network including a parameter group of the second parameters and the fourth parameters, the inference apparatus comprising` processing circuitry configured to:
[0118]:
a plurality of patterns of learning models such as a learning model according to an examination site of the subject
[0093]:
in the feature quantity space (feature quantity map), a relationship in image data among teacher data 1002, image data before conversion 1001 and after conversion 1003 from the feature quantity B to the feature quantity C is analyzed. For example, a feature quantity space error reduction is particularly weighted to the distance between a patch image (group) 1402 having a deep relation to determination of the presence or absence of a disease in a pathological image and an MRI image (group) 1401 corresponding thereto.
In [0104]:
the feature quantity conversion unit 234 uses the feature quantity conversion model 234M (FIG. 6) to convert the feature quantity B into the feature quantity C (S5). The identification unit 235 uses the identification model 235M to calculate a parameter value for predicting
- input processing target data to the first network and generate a first processing result.
[0118]:
a plurality of patterns of learning models such as a learning model according to an examination site of the subject
[0041]:
As illustrated in FIG. 1, a medical imaging apparatus 10 according to the present embodiment includes an imaging unit 100 that collects an image signal necessary for image reconstruction from a subject and an image processing unit 200 that performs image processing of the subject imaged by the imaging unit 100. The medical imaging apparatus 10 further includes an input unit 110.
[0013]:
Specifically, a medical imaging apparatus includes an imaging unit that collects an image signal of an inspection target, and an image processing unit that generates first image data from the image signal and performs image processing of the first image data. The image processing unit includes a feature quantity extraction unit that extracts a first feature quantity from the first image data, a feature quantity abstraction unit that extracts (abstracts) a more important second feature quantity from the first feature quantity, a feature quantity conversion unit that converts the second feature quantity into a third feature quantity extracted by second image data different from the first image data, and an identification unit that uses the converted third feature quantity to calculate a predetermined parameter value, and performs prediction.
- input processing target data to the second network and generate a second processing result.
[0118]:
a plurality of patterns of learning models such as a learning model according to an examination site of the subject
[0041]:
As illustrated in FIG. 1, a medical imaging apparatus 10 according to the present embodiment includes an imaging unit 100 that collects an image signal necessary for image reconstruction from a subject and an image processing unit 200 that performs image processing of the subject imaged by the imaging unit 100. The medical imaging apparatus 10 further includes an input unit 110.
[0013]:
Specifically, a medical imaging apparatus includes an imaging unit that collects an image signal of an inspection target, and an image processing unit that generates first image data from the image signal and performs image processing of the first image data. The image processing unit includes a feature quantity extraction unit that extracts a first feature quantity from the first image data, a feature quantity abstraction unit that extracts (abstracts) a more important second feature quantity from the first feature quantity, a feature quantity conversion unit that converts the second feature quantity into a third feature quantity extracted by second image data different from the first image data, and an identification unit that uses the converted third feature quantity to calculate a predetermined parameter value, and performs prediction.
- and calculate at least one of a weighted average of the first processing result and the second processing result and reliability based on a difference between the first processing result and the second processing result.
[0095] :
in FIG. 15 illustrating a feature quantity space, an error function in which weights are given to a distance between medically important teacher data A1 and an output B1 and a distance between next important teacher data group (data set S) Ai and an output Bi is set. When a conversion function to the feature quantity space ε is set to g and the center of gravity (average value of coordinates of each piece of data) on the space ε is represented by C, the error function is represented by the following Formula (6).
PNG
media_image12.png
107
363
media_image12.png
Greyscale
[0096]:
Here, α, β, and γ are weighting factors, for example, α=0.5, β=0.4, and γ=0.1.
[0097]:
By using the error function as described above, it is possible to reduce the error of the feature quantity conversion model or the identification model and realize a more accurate predictive model. Alternatively, the error functions (2) and (5) may be combined and weighted to form an error function represented by the following Formula (7).
PNG
media_image13.png
26
377
media_image13.png
Greyscale
[0098]:
Here, w1 and w2 are weighting factors (for example, w1=0.5, w2=0.5).
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 TIRUMALE KRISHNASWAMY RAMESH whose telephone number is (571)272-4605. The examiner can normally be reached by phone.
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 B Zhen can be reached on phone (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.
/TIRUMALE K RAMESH/
Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
