Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 06/23/2026 has been entered.
Response to Arguments
Applicant’s argument filed 06/23/2026 have been fully considered but they are not persuasive.
Applicant’s Argument: On pages 8-12 of Applicant’s response, applicant states that the claims do not fall within any enumerated grouping of abstract ideas. The amended claims are not directed towards mental processes because the diffusion and denoising operations cannot be performed mentally. The amended claims are also similar to Example 39 of training a machine learning model for a specific purpose.
Examiner’s Response: Applicant’s argument is not persuasive. Claim 4 recites performing encoding operations on a first data point, performing one or more diffusion operations on the latent variables, and performing denoising operations on the first set of values. Under the broadest reasonable interpretation, the first data point can encompass any data such as a set of numbers. The process of encoding numerical data includes converting numbers into binary form and thus, encoding a first data point can be performed in the human mind using pen and paper. The operations of diffusion and denoising are adding and removing noise from the data. The process of adding additional bits to the encoded numerical data can be performed in the human mind using pen and paper. The process of removing specific bits from the noisy numerical data can also be performed in the human mind using pen and paper. Therefore, the recited claims amount to an abstract idea of a mental process. Training a neural network is a generic computer process and step of encoding, diffusion, and denoising are mere instructions for using a computer as a tool to perform a mental process. The use of a computer or other machinery in its ordinary capacity does not integrate a judicial exception into a practical application or provide significantly more.
In Example 39, the claim does not recite any of the judicial exceptions under Subject Matter Eligibility Analysis Step 2A Prong 1. Thus, Example 39 is not applicable to the claims of the invention as the amended claims do recite judicial exceptions.
Applicant’s Argument: On pages 12-14 of Applicant’s response, applicant states that the claims are directed to a technological improvement of the training of a generative model using a score-based generative model to generate new data. Specifically, the improvement is recited as performing one or more encoding operations on a first data point, performing one or more diffusion operations on the first set of latent variables, performing one or more denoising operations on the first set of values, and generating a trained generative model.
Examiner’s Response: Applicant’s argument is not persuasive. During examination, the examiner should analyze the "improvements" consideration by evaluating the specification and the claims to ensure that a technical explanation of the asserted improvement is present in the specification, and that the claim reflects the asserted improvement (see MPEP §2106.05(a)). The MPEP (§2106.05(a)(II)) also warns, “it is important to keep in mind that an improvement in the abstract idea itself (e.g. a recited fundamental economic concept) is not an improvement in technology.” Here, the alleged improvement in the form of “performing one or more encoding operations on a first data point, performing one or more diffusion operations on the first set of latent variables, and performing one or more denoising operations on the first set of values” is an improvement to the abstract idea of a mental process that can be performed in the human mind. Additionally, the claim limitation “generating a trained generative model based on the one or more losses” is a generic computer process and the use of a computer or other machinery in its ordinary capacity does not integrate a judicial exception into a practical application or provide significantly more.
An important consideration in determining whether a claim improves technology is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome, as opposed to merely claiming the idea of a solution or outcome (see MPEP 2106.05(a)). The amended claims do not provide sufficient details to describe any technological improvement. If the specifications explicitly set forth an improvement but in a conclusory manner (see MPEP 2106.04(d)(1): a bare assertion of an improvement without the detail necessary to be apparent to a person of ordinary skill in the art), the examiner should not determine the claim improves technology.
Applicant’s Argument: On pages 14-17 of Applicant’s response, applicant states that Karatsiolis would have to disclose that the classifier performs diffusion operations on the latent representation to teach the amended claims. Karatsiolis does not disclose the classifier performs any diffusion operations and fails to disclose the feature mapping is associated with a base distribution of the classifier. Applicant also argues that Song does not receive encoded latent variables value from an encoder network and performing diffusion operations on the encoded values. The combination of the cited references fails to teach the amended claims.
Examiner’s Response: Applicant’s argument is not persuasive. Applicant’s arguments with respect to claim 1 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. The amendments to claim 1 change the scope of the invention. The combination of the cited references still teaches the claim as a whole.
In response to applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).
The claimed invention is rejected under 35 U.S.C. 103 as being unpatentable over Karatsiolis in view of Song. The claimed invention discloses an autoencoder with an embedded score-based generative model. Karatsiolis (pg. 3, Figure 1) teaches the structure of the claimed invention because Karatsiolis discloses an autoencoder with an embedded classifier. Karatsiolis (pg. 3, Section B, par. 1-2) discloses that the VAE applies an isotropic Gaussian distribution on the latent variable and the latent distribution is processed by the classifier. The latent variables are processed by the hidden layers of the classifier to generate a latent distribution. Under the broadest reasonable interpretation, the output of the classifier is associated with the base distribution of the VAE. Karatsiolis in combination of Song is used to teach the claim invention because the classifier in Karatsiolis can be replaced by the score-based generative model taught in Song.
Karatsiolis in view of Song teaches the amended claims. Karatsiolis (pg. 3, Fig, 2) discloses the input data is encoded into a latent representation Z1 and the latent representation is passed onto the classifier. Karatsiolis (pg. 5, Section V, par. 1) discloses the classifier consists of multiple hidden layers that perform various operations such as activation functions and feature mapping. The score-based generative model from Song can be incorporated into the structure disclosed by Karatsiolis to replace the classifier and the score-based generative model from Song can be used to teach the specific operations performed on the encoded data. The score-based generative model from Song (pg 3-4, Section 3.1 & 3.2) performs a diffusion process and a reverse diffusion process as disclosed by the amended claims.
Examiner has provided Huang as a supplementary reference. Huang (pg. 6, Section 5 & 6) discloses the integration of a infinitely deep hierarchical VAE into the score-based generative modeling framework.
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 4-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Subject Matter Eligibility Analysis Step 1:
Claims 4-13 recite a process (“A computer-implemented method for training a generative model, the method comprising”), one of the four statutory categories of patentable subject matter. Claims 14-20 recite an article of manufacture (“One or more non-transitory computer readable media …”), one of the four statutory categories of patentable subject matter.
Regarding Claim 4:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“performing, a first data point included in a training dataset to generate a first set of latent variable values” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
“performing, a first set of values associated with a base distribution for the score-based generative model” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
“performing one or more denoising operations , a second set of latent variable values associated with a continuous latent space ” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
“performing one or more additional operations to convert the second set of latent variable values into a second data point” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
“computing one or more losses based on the first data point and the second data point” (a mathematical calculation; see par. 135 in Specification, computing a reconstruction loss)
Claim 4 therefore recites an abstract idea.
Subject Matter Eligibility Analysis Step 2A Prong 2:
“performing, via an encoder network …”; “performing, via a score-based generative mode …” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“performing one or more denoising operations via the score-based generative model the encoder network” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based generative model” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
The additional elements as disclosed above alone or in combination do not integrate the judicial exception into practical application as they are mere insignificant extra solution activity in combination of generic computer functions being implemented with generic computer elements in a high level of generality to perform the disclosed abstract idea above. Therefore, Claim 4 is directed to the abstract idea.
Subject Matter Eligibility Analysis Step 2B:
“performing, via an encoder network …”; “performing, via a score-based generative mode …” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“performing one or more denoising operations via the score-based generative model generative model,the encoder network” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based generative model” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
The additional elements as disclosed above alone or in combination do not recite significantly more than the abstract idea itself as they are mere insignificant extra solution activity in combination of generic computer functions being implemented with generic computer elements in a high level of generality to perform the disclosed abstract idea above. Therefore, Claim 4 is subject-matter ineligible.
Regarding Claim 14:
The claim recites an article of manufacture that performs the method as described in claim 4. Therefore, claim 14 is rejected for the same reasons as disclosed for claim 4. The limitations for additional elements of claim 14 are analyzed below.
Subject Matter Eligibility Analysis Step 2A Prong 1:
Please see Step 2A Prong 1 analysis of claim 4
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“One or more non-transitory computer readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 5:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“performing one or more encoding operations the first set of latent variable values” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
“performing the one or more diffusion operations to convert the first set of latent variable values into the first set of values” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“performing one or more encoding operations via the encoder network” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 6:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein performing the one or more additional operations comprises applying a decoder neural network to the second set of latent variable values to produce the second data point” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claims 7 and 17:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“wherein computing the one or more losses comprises computing a cross-entropy loss associated with a first distribution of the second set of latent variable values the first set of latent variable values ” (a mathematical calculation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“the encoder network based on the training dataset” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 8:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“wherein computing the cross-entropy loss comprises sampling from a proposal distribution associated with a loss weighting included in the cross-entropy loss” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B: None
Regarding Claim 9:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the loss weighting comprises a diffusion coefficient associated with a diffusion process between the continuous latent space and the base distribution” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 10:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the cross-entropy loss comprises at least one of a first loss weighting associated with the encoder network and a second loss weighting associated with the score-based generative model” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
Regarding Claim 11:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein generating the trained generative model comprises” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claims 12 and 19:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“computing a reconstruction loss associated with the first data point and the second data point” (a mathematical calculation)
“computing a negative encoder entropy loss associated with the first set of latent variable values ” (a mathematical calculation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 13:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“third set of latent variable values in order to generate a new data point that is not included in the training dataset” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein, in operation, the trained generative model converts” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 15:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“the first set of latent variable values and second set of latent variable values into the second data point” (a mental process that can be performed in the human mind with the aid of pen and paper, i.e. evaluation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the instructions further cause the one or more processors to perform the step of generating a pre-trained encoder neural network and a pre-trained decoder neural network included in the score-based generative model based on a standard Normal prior, wherein the pre-trained encoder neural network converts ” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 16:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein generating the trained generative model comprises performing end-to-end training of the pre-trained encoder neural network, the pre-trained decoder neural network, and the score-based generative model based on the one or more losses” (mere instructions to apply the exception using a generic computer component - see MPEP 2106.05(f))
Regarding Claim 18:
Subject Matter Eligibility Analysis Step 2A Prong 1:
“wherein computing the cross-entropy loss comprises computing the cross-entropy loss based on a geometric variance associated with the one or more denoising operations” (a mathematical calculation)
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B: None
Regarding Claim 20:
Subject Matter Eligibility Analysis Step 2A Prong 1: None
Subject Matter Eligibility Analysis Step 2A Prong 2 & 2B:
“wherein the score-based generative model comprises a set of residual network blocks” (merely specifies a particular technological environment in which the abstract idea is to take place, ie. a field of use, and thus does not integrate the abstract idea into a practical application nor cannot provide significantly more than the abstract idea itself - see MPEP 2106.05(h))
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 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Karatsiolis, "Conditional Generative Denoising Autoencoder" in view of Song, "Score-Based Generative Modeling Through Stochastic Differential Equations".
Regarding claim 1, Karatsiolis teaches:
“A computer-implemented method for training a generative model, the method comprising” (abstract, A generative denoising autoencoder model (generative model) is trained to produce better data samples from a set of images.)
“performing, via an encoder network, one or more encoding operations on a training image included in a training dataset to generate a first set of latent variable values” ([pg. 3, section III, par. 1-2; pg. 5, section V, par. 1; pg. 3, Figure 2], The encoder structure generates the latent representation Z1 from the input data, which are image.)
“performing, via a score-based [discriminative] model, one or more a first set of values associated with a base distribution for the score-based ” ([pg. 3-4, section III, par. 1-2 & 6; pg. 5-6, section V, par. 1; pg. 3, Figure 1], The output of the encoder is sent to the classifier (scored-based model). The latent representation contributed by the classifier is comprised of the outputs of its hidden layers. The classifier generates internal representations using its hidden layers based on the latent representation received by the encoder. In some embodiment, the classifier may consist of 3 hidden layers. Thus, the input latent representation passes through one or more hidden layer to generate feature mappings.)
“performing one or more denoising operations via the score-based second set of latent variable values associated with a continuous latent space learned by the encoder network” ([pg. 3, section III, par. 1-2; pg. 5, section IV, par. 4; pg. 5-6, section V, par. 1; pg. 3, Figure 2], The classifier performs feature mapping (denoising operations) to generate Z2 (second set of latent variable values) that encapsulate discriminative information. The output of the classifier is latent representation. In some embodiment, the third hidden layer receives output from the previous hidden layer to generate a new feature mapping. The latent distribution is formed by merging the contribution of the encoder with the internal representations of the classifier.)
“performing one or more additional operations to convert the second set of latent variable values into an output image” ([pg. 5, section IV, par. 4; pg. 3, Figure 2], The joint distribution is input to the decoder and the output is a generated image by the model that resembles the input image.)
“computing one or more losses based on the training image and the output image” ([pg. 4, col. 1, par. 2-4], The two objective functions are shown in page 4 of the reference. The unsupervised objective function calculates a loss between the input data, x1 and the output image from the decoder.)
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based ” ([pg. 4, col. 1, par. 2-4; pg. 3, Figure 1], The unsupervised objective function is used to train the generative denoising autoencoder model. The generative denoising autoencoder model contains an autoencoder and an embedded classifier, which has the score-based learning.)
Karatsiolis does not explicitly disclose an implementation of a score-based generative model to perform the claimed limitation “performing, via a score-based generative model, one or more diffusion operations” and “performing one or more denoising operations via the score-based generative model”. However, Song discloses in the same field of endeavor:
“performing, via a score-based generative model, one or more diffusion operations on the first set of latent variable values a first set of values associated with a base distribution for the score-based generative model” ([pg. 3-4, section 3.1, par. 1-2; pg. 8-9, section 5, par. 3; pg. 4, Figure 2], A forward SDE maps input data to a noise distribution. The prior distribution of the generative model is constructed based on the data distribution using a diffusion process. Under the broadest reasonable interpretation, the intermediate data points along the trajectory of a forward stochastic differential equation (SDE) are latent variables. In one embodiment, the score-based generative model may be a time-dependent classifier.)
“performing one or more denoising operations via the score-based generative model on the first set of values to output, from the score-based generative model, a second set of latent variable values associated with a continuous latent space ” ([pg. 3-4, section 3.1, par. 1-2; pg. 4, section 3.2, par. 1; pg. 4, Figure 2], The generative model performs a reverse-time SDE (denoising operations) on the samples associated with the prior distribution. Data is mapped to an unstructured prior distribution with forward SDE and intermediate latent variables can be sampled from the prior distribution by reversing the SDE process. The score-based generative model performs the reverse diffusion process to generate an output.)
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based generative model” ([pg. 4-5, section 3.3, par. 1-3], The score (one or more losses) is calculated and used to train the score-based generative model.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “performing, via a score-based generative model, one or more diffusion operations” and “performing one or more denoising operations via the score-based generative model” from Song into the teaching of Karatsiolis. The score-based generative model from Song can replace the classifier from Karatsiolis. Doing so can allow for new sampling procedures and new modeling capabilities by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 2, Karatsiolis teaches:
“wherein the trained generative model further includes a decoder neural network that converts the second set of latent variable values into the output image.” ([pg. 5, section IV, par. 4; pg. 3, Figure 2], The joint distribution is input to the decoder and the output is a generated image by the model that resembles the input image.)
Regarding claim 3, Karatsiolis teaches:
“wherein, in operation, the trained generative model converts a second set of values associated with the base distribution into a third set of latent variable values in order to generate a new image that is not included in the training dataset.” ([pg. 5, section IV, par. 4-5; pg. 3, Figure 2], The flow of the input data through the structural components shown in Figure 2 is an iterative process. The generated output is modified and used as input to the generative model. The new input is propagated to the network to generate a new joint distribution, which the decoder structure uses to generate a new image.)
Regarding claim 4, Karatsiolis teaches:
“A computer-implemented method for training a generative model, the method comprising” (abstract, A generative denoising autoencoder model (generative model) is trained to produce better data samples from a set of images.)
“performing, via an encoder network, one or more encoding operations on a first data point included in a training dataset to generate a first set of latent variable values” ([pg. 3, section III, par. 1-2; pg. 5, section V, par. 1; pg. 3, Figure 2], The encoder structure generates the latent representation Z1 from the input data.)
“performing, via a score-based [discriminative] model, one or more a first set of values associated with a base distribution for the score-based ” ([pg. 3-4, section III, par. 1-2 & 6; pg. 5-6, section V, par. 1; pg. 3, Figure 1], The output of the encoder is sent to the classifier (scored-based model). The latent representation contributed by the classifier is comprised of the outputs of its hidden layers. The classifier generates internal representations using its hidden layers based on the latent representation received by the encoder. In some embodiment, the classifier may consist of 3 hidden layers. Thus, the input latent representation passes through one or more hidden layer to generate feature mappings.)
“performing one or more denoising operations via the score-based second set of latent variable values associated with a continuous latent space learned by the encoder network” ([pg. 3, section III, par. 1-2; pg. 5, section IV, par. 4; pg. 5-6, section V, par. 1; pg. 3, Figure 2], The classifier performs feature mapping (denoising operations) to generate Z2 (second set of latent variable values) that encapsulate discriminative information. The output of the classifier is latent representation. In some embodiment, the third hidden layer receives output from the previous hidden layer to generate a new feature mapping. The latent distribution is formed by merging the contribution of the encoder with the internal representations of the classifier.)
“performing one or more additional operations to convert the second set of latent variable values into a second data point” ([pg. 5, section IV, par. 1-4; pg. 3, Figure 2], The joint distribution is input to the decoder and the output is a generated image by the model that resembles the input image.)
“computing one or more losses based on the first data point and the second data point” ([pg. 4, col. 1, par. 2-4], The two objective functions are shown in page 4 of the reference. The unsupervised objective function calculates a loss between the input data, x1 and the output data from the decoder.)
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based ” ([pg. 4, col. 1, par. 2-4], The unsupervised objective function is used to train the generative denoising autoencoder model. The generative denoising autoencoder model contains an autoencoder and an embedded classifier, which has the score-based learning.)
Karatsiolis does not explicitly disclose an implementation of a score-based generative model to perform the claimed limitation “performing, via a score-based generative model, one or more diffusion operations” and “performing one or more denoising operations via the score-based generative model”. However, Song discloses in the same field of endeavor:
“performing, via a score-based generative model, one or more diffusion operations on the first set of latent variable values a first set of values associated with a base distribution for the score-based generative model” ([pg. 3-4, section 3.1, par. 1-2; pg. 8-9, section 5, par. 3; pg. 4, Figure 2], A forward SDE maps input data to a noise distribution. The prior distribution of the generative model is constructed based on the data distribution using a diffusion process. Under the broadest reasonable interpretation, the intermediate data points along the trajectory of a forward stochastic differential equation (SDE) are latent variables. In one embodiment, the score-based generative model may be a time-dependent classifier.)
“performing one or more denoising operations via the score-based generative model on the first set of values to output, from the score-based generative model, a second set of latent variable values associated with a continuous latent space ” ([pg. 3-4, section 3.1, par. 1-2; pg. 4, section 3.2, par. 1; pg. 4, Figure 2], The generative model performs a reverse-time SDE (denoising operations) on the samples associated with the prior distribution. Data is mapped to an unstructured prior distribution with forward SDE and intermediate latent variables can be sampled from the prior distribution by reversing the SDE process. The score-based generative model performs the reverse diffusion process to generate an output.)
“generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based generative model” ([pg. 4-5, section 3.3, par. 1-3], The score (one or more losses) is calculated and used to train the score-based generative model.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “performing, via a score-based generative model, one or more diffusion operations” and “performing one or more denoising operations via the score-based generative model” from Song into the teaching of Karatsiolis. The score-based generative model from Song can replace the classifier from Karatsiolis. Doing so can allow for new sampling procedures and new modeling capabilities by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 14:
Claim 14 recites an article of manufacture (“One or more non-transitory computer readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of”) that performs the same process as described in Claim 4. Therefore claim 14 is rejected under the same reasons mention for claim 4. The additional elements of claim 14 is addressed below by Karatsiolis:
“One or more non-transitory computer readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of” ([pg. 3, section III, par. 1; Figure 1], The generative model is shown in Figure 1. It is implied that the architecture is created on a computer system that consists of processors and memory for running the process described by the reference.)
Regarding claim 5, Karatsiolis teaches:
“performing one or more encoding operations via the encoder network to convert the first data point into the first set of latent variable values.” ([pg. 3, section III, par. 1-2; pg. 5, section V, par. 1; pg. 3, Figure 2], The encoder structure generates the latent representation Z1 from the input data.)
“performing the one or more first set of latent variable values into the first set of values” ([pg. 3, section III, par. 1-2; pg. 5, section V, par. 1; Figure 2], The classifier receives the latent representation Z1 and processes the latent representation using one or more hidden layers of the classifier.)
Karatsiolis does not explicitly disclose an implementation of “performing the one or more diffusion operations to convert the first set of latent variable values into the first set of values”. However, Song discloses in the same field of endeavor:
“performing the one or more diffusion operations to convert the first set of latent variable values into the first set of values” ([pg. 3, section 3.1, par. 1; pg. 4, Figure 2], The generative model performs a diffusion process that consists of perturbing the input data with noise.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “performing the one or more diffusion operations to convert the first set of latent variable values into the first set of values” from Song into the teaching of Karatsiolis. Doing so can allow for new sampling procedures and new modeling capabilities by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 6, Karatsiolis teaches:
“wherein performing the one or more additional operations comprises applying a decoder neural network to the second set of latent variable values to produce the second data point” ([pg. 5, section IV, par. 1-4; Figure 2], The jkoint latent distribution is provided to the decoder to produce an output.)
Regarding claims 7 and 17, Karatsiolis teaches:
“wherein computing the one or more losses comprises computing a cross-entropy loss associated with a first distribution of the second set of latent variable values generated by the score-based the first set of latent variable values generated by the encoder network based on the training dataset” ([pg. 4, col. 1, par. 2-4], The two objective functions are shown in page 4 of the reference. The supervised objective function computes a cross-entropy loss that is related to the classifier and the encoder. In the equation, Z1 and QΘ are related to the encoder and Pw is related to the classifier. The supervised objective function is a calculation of the loss that is associated with the classifier and encoder.)
Karatsiolis does not explicitly disclose an implementation of a score-based generative model. However, Song discloses in the same field of endeavor:
“wherein computing the one or more losses comprises computing a cross-entropy loss associated with a first distribution of the second set of latent variable values generated by the score-based generative model The score (one or more losses) is calculated using Equation 7. The score-based generative model can be substitute for the classifier)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of a score-based generative model from Song into the teaching of Karatsiolis. Doing so can allow for new sampling procedures and new modeling capabilities by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 8, Karatsiolis teaches:
“wherein computing the cross-entropy loss comprises sampling from a proposal distribution associated with In determining the learning gradient information, the cross-entropy loss is calculated with respect to the sampling of an input data from a uniform distribution.)
Karatsiolis does not explicitly disclose an implementation of “sampling from a proposal distribution associated with a loss weighting included in the cross-entropy loss”. However, Song discloses in the same field of endeavor:
“wherein computing the cross-entropy loss comprises sampling from a proposal distribution associated with a loss weighting included in the cross-entropy loss” ([pg. 4-5, section 3.3, par. 1-3, Equation 7], Equation 7 shows the calculation of the cross-entropy loss function and λ is a loss weighting parameter.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of a score-based generative model from Song into the teaching of Karatsiolis. Doing so can improve the performance of the generative model by incorporating architectural changes that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 9, Karatsiolis does not explicitly disclose an implementation of “wherein the loss weighting comprises a diffusion coefficient associated with a diffusion process between the continuous latent space and the base distribution”. However, Song discloses in the same field of endeavor:
“wherein the loss weighting comprises a diffusion coefficient associated with a diffusion process between the continuous latent space and the base distribution” ([pg. 3-4, section 3.1, par. 1, Equation 5-7], Equation 5 shows the calculation relating the data distribution and the prior distribution. The function g is the diffusion coefficient of the data.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of a “wherein the loss weighting comprises a diffusion coefficient associated with a diffusion process between the continuous latent space and the base distribution” from Song into the teaching of Karatsiolis. Doing so can improve the performance of the generative model by incorporating architectural changes that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 10, Karatsiolis teaches:
“wherein the cross-entropy loss comprises at least one of a first loss weighting associated with the encoder network and a second loss weighting associated with the score-based The two objective functions are shown in page 4 of the reference. The supervised and unsupervised objective functions compute a cross-entropy loss that is related to the classifier and the encoder. The loss is based on backpropagating the gradient information back to the encoder and classifier to update their weights during a series of mini-batch.)
Karatsiolis does not explicitly disclose an implementation of a score-based generative model. However, Song discloses in the same field of endeavor:
“wherein the cross-entropy loss comprises at least Equation 7 shows the calculation of the cross-entropy loss function and λ is a loss weighting parameter. A score-based generative model can be substitute for the classifier)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of a score-based generative model from Song into the teaching of Karatsiolis. Doing so can allow for new sampling procedures to generate image samples of higher quality by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract)
Regarding claim 11, Karatsiolis teaches:
“wherein generating the trained generative model comprises updating a plurality of parameters associated with the score-based The parameters of the objective function can be updated during the training procedure of the generative model. During the training with unlabeled and labeled data, the parameters of the encoder and classifier are updated using gradient information.)
Karatsiolis does not explicitly disclose an implementation of a score-based generative model. However, Song discloses in the same field of endeavor:
“wherein generating the trained generative model comprises updating a plurality of parameters associated with the score-based generative model The score of the score-based generative model is calculated using Equation 7. A score-based generative model can be substitute for the classifier.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of a score-based generative model from Song into the teaching of Karatsiolis. Doing so can allow for new sampling procedures to generate image samples of higher quality by implementing a framework that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract)
Regarding claims 12 and 19, Karatsiolis teaches:
“computing a reconstruction loss associated with the first data point and the second data point” ([pg. 4, col. 1, par. 2-4], The unsupervised objective function calculates a loss between the input data, x1 and the output data from the decoder.)
“computing a negative encoder entropy loss associated with the first set of latent variable values generated by the encoder network based on the training dataset” ([pg. 7, col. 1, par. 2], During a specific instance of the iterative process of the generation model, negative gradient information (negative encoder entropy loss) may be provided back into the model to generate a second set of Z1 latent representations to specifically remove certain image properties in the generated image.)
Regarding claim 13, Karatsiolis teaches:
“wherein, in operation, the trained generative model converts a second set of values associated with the base distribution into a third set of latent variable values in order to generate a new data point that is not included in the training dataset” ([pg. 5, section IV, par. 4-5; Figure 2 and 7], The flow of the input data through the structural components shown in Figure 2 is an iterative process. The generated output is modified and used as input to the generative model. The new input is propagated to the network to generate a new joint distribution, which the decoder structure uses to generate a new image. Figure 7 shows the iterative process of the generative model to start with an initial image and generate a series of new images until a target image is produced.)
Regarding claim 15, Karatsiolis in view of Song teaches:
“wherein the instructions further cause the one or more processors to perform the step of generating a pre-trained encoder neural network and a pre-trained decoder neural network included in the score-based Table I shows the process of training the generative denoising autoencoder. Step 4 of the training procedure involves unsupervised learning to train and update the encoder and decoder using the generated gradient information. During the unsupervised learning mini-batch, the parameters of the classifier are not updated. VAE apply a Gaussian distribution on the latent variable and maximize the variational lower bound of the model. Song (pg. 4, Figure 2) further discloses the score-based generative model through SDEs.)
Regarding claim 16, Karatsiolis in view of Song teaches:
“wherein generating the trained generative model comprises performing end-to-end training of the pre-trained encoder neural network, the pre-trained decoder neural network, and the score-based Table I shows the process of training the encoder, decoder, and classifier. A gradient descent method (one or more losses) is used to train the 3 components of the system. Song (pg. 4, Figure 2) further discloses the score-based generative model through SDEs.)
Regarding claim 18, Karatsiolis does not explicitly disclose an implementation of “wherein computing the cross-entropy loss comprises computing the cross-entropy loss based on a geometric variance associated with the one or more denoising operations”. However, Song discloses in the same field of endeavor:
“wherein computing the cross-entropy loss comprises computing the cross-entropy loss based on a geometric variance associated with the one or more denoising operations” ([pg. 15, Appendix C, par. 1], Appendix A-C provides details on how general SDEs can be applied to the framework discussed in the reference. An implementation where a geometric sequence relating the injecting noise into the data can be applied to the diffusion process.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein computing the cross-entropy loss comprises computing the cross-entropy loss based on a geometric variance associated with the one or more denoising operations” from Song into the teaching of Karatsiolis. Doing so can improve the performance of the generative model by incorporating architectural changes that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Regarding claim 20, Karatsiolis does not explicitly disclose an implementation of “wherein the score-based generative model comprises a set of residual network blocks”. However, Song discloses in the same field of endeavor:
“wherein the score-based generative model comprises a set of residual network blocks” ([pg. 26, Appendix H.2, par. 1-4], Appendix provides details on how improvements can be applied to the framework discussed in the reference. The score-based generative model has the residual blocks modified for performance improvements. The residual blocks are used for feature mapping in the process.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein the score-based generative model comprises a set of residual network blocks” from Song into the teaching of Karatsiolis. Doing so can improve the performance of the generative model by incorporating architectural changes that encapsulates score-based generative modeling and diffusion probabilistic modeling (Song, abstract).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Huang (“A Variational Perspective on Diffusion-Based Generative Models and Score Matching”) is included in this Office Action. Huang (abstract) is an extension to the Song reference and implements a variational framework for likelihood estimation. Huang (pg. 6, section 5, par. 1-2) discloses the use of an infinitely deep hierarchical VAE into the score matching diffusion generative model.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GARY MAC whose telephone number is (703)756-1517. The examiner can normally be reached Monday - Friday 8:00 AM - 5:00 PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Abdullah Kawsar can be reached on (571) 270-3169. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/GARY MAC/Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
Prosecution Insights