DETAILED ACTION
This action is responsive to the application filed on 2/21/2024. Claims 1-25 are pending in the case. Claims 1, 13 and 25 are independent claims.
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 .
Drawings
The drawings are objected to because of the following:
In Fig. 7, Ref. No. 700 is defined as the entire manufacturing system, but in the drawing, it appears to point to a specific part of the manufacturing system. In accordance with reference numbers 100, 200, 300, 400, 500, and 600, Ref. No. 700 should point to the manufacturing system as a whole and not a specific part of the system. If the arrow and location of Ref. No. 700 is accurate in the drawing, then the examiner requires that the specification be amended to more reasonably explain the location of Ref. No. 700 and its connection to the manufacturing machine 702 and the parts of the system described in Fig. 6 and carried over to Fig. 7
Corrected drawing sheets in compliance with 37 CFR 1.121(d) are required in reply to the Office action to avoid abandonment of the application. Any amended replacement drawing sheet should include all of the figures appearing on the immediate prior version of the sheet, even if only one figure is being amended. The figure or figure number of an amended drawing should not be labeled as “amended.” If a drawing figure is to be canceled, the appropriate figure must be removed from the replacement sheet, and where necessary, the remaining figures must be renumbered and appropriate changes made to the brief description of the several views of the drawings for consistency. Additional replacement sheets may be necessary to show the renumbering of the remaining figures. Each drawing sheet submitted after the filing date of an application must be labeled in the top margin as either “Replacement Sheet” or “New Sheet” pursuant to 37 CFR 1.121(d). If the changes are not accepted by the examiner, the applicant will be notified and informed of any required corrective action in the next Office action. The objection to the drawings will not be held in abeyance.
Abstract
The abstract of the disclosure is objected to because of the following:
The abstract should be able to convey a brief and comprehensible description of the claimed invention and should therefore be free from issues regarding antecedent basis. Referring to elements like, “the generative model,” before they have been introduced creates issues with antecedent basis and should be amended; see the 112B rejections below for elements in the claims and abstract alike that lack antecedent basis
In the last line, the abstract reads, “a reverse of the reversible Heun SDE solver is used,” and should read “a reverse algorithm of the Reversible-Heun SDE solver is used” in order to maintain consistency in the titles of elements and ensure that the reader understand the reverse is an algorithm
A corrected abstract of the disclosure is required and must be presented on a separate sheet, apart from any other text. See MPEP § 608.01(b).
Specification
The disclosure is objected to because of the following informalities:
The sections titled “Technical Field” and “Background” should be in one section titled “Background of the Invention” or something of the like and the currently titled “Background” section should be renamed to “Description of Related Art” or something of the like
In paragraph [0086], line 2, the computer-controlled machine is referred to using Ref. No. 600 and should be corrected to Ref. No. 602
In paragraph [0088], line 5, “process each sensor signal 618 to product each input signal x” should read “process each sensor signal 618 to produce each input signal x”
In paragraph [0089], line 2, “models such ash those…” should read “models such as those…”
In paragraph [0090], line 2, “related actuator control command 20” should read “related actuator control commands 620,” which pluralizes commands as in the rest of the specification and corrects the Ref. No. used to refer to such commands
In paragraph [0092], line 3, Ref. No. 624 is referred to as “The classifier 624” despite being introduced as “ML Processing.” Although it is understood that ML Processing can be configured to classify, it is important to remain consistent in the matching of an element titles to reference numbers. The examiner suggests replacing “The classifier 624…” with “ML Processing 624 (configured to classify)…” to correct this issue
Appropriate correction is required.
Claim Objections
Claims 1-25 objected to because of the following informalities:
Claims 1, 2, 13, 14, and 25 should all refer to the ‘clean example’ or ‘clean image’ in the same way, there is no reason to change the wording of the element when they are all meant to represent the same thing. The examiner suggests using ‘clean image’ for all such references to this element as it is more specific. Additionally, all of the references to this element should be either ‘sampled clean image’ or ‘clean image,’ (other than the initial introduction of “sampling the clean image” in the sampling step of the independent claims) so as to remain consistent in the title of the element.
Claims 1, 13, and 25 recite “sampling a clean example to be learned” or “sampling a clean image to be learned.” The examiner asserts that the addition of the phrase “to be learned” is an inherit condition of data during training of a model; the phrase should be removed to prevent misconstruction of the limitation. The examiner suggests rephrasing of the limitation to something along the lines of: “sampling a clean example from an input set of training data used to train the generative model”
Claims 1, 13, and 25 have irregular spacing when reciting the “Schrodinger Bridge Forward-Backward Stochastic Differential Equations (SB-FBSDEs).” The spacing between words should remain consistent throughout the entire disclosure, including the set of claims.
Claims 1 and 3 recite “…the generative modeling problem.” This limitation can be reasonably understood to relate to the training that is taking place, but the examiner requires that the phrasing of the limitation is modified to prevent misconstruction of the claim. One could read the limitation and be confused as to what “the generative modeling problem” refers to. Some suggestions for rephrasing include: “…computed for the generative model” or “…computed for the generative model training” to make the limitation clearer, see claims 13 and 25. Note that this is not a rejection under 112B because, given the specification, one could interpret what the “problem” refers to.
Claims 1, 13, and 25 recite “…computed for the generative modeling problem, producing predicted output values” and should read “…computed for the generative modeling problem and produce predicted output values” or something of the like. The recited correction will make the limitation more grammatically correct and comprehensible in terms of its intended meaning given the specification.
Claims 1, 13, and 25 recite “computing a loss function to compares true values…” and should read “computing a loss function to compare true values…,” fixing the grammatical error
Claims 1, 13, and 25 recite “using a reverse algorithm of the RH SDE solver, solving the stochastic adjoint SDE to compute…” and should read “using a reverse algorithm of the RH SDE solver to solve the stochastic adjoint SDE, to compute…” The recited correction will make the limitation more grammatically correct and comprehensible in terms of its intended meaning given the specification.
Claims 4 refers to the generative model as “the deep generative model” and should be amended to remove the word “deep.” The examiner asserts that it is important to remain consistent in the titling of elements to prevent misconstruction of the limitations.
Claims 6 and 18 recite “repeating the training a plurality of times” and should potentially read “repeating the training until convergence.” A plurality of times implies 2 or more times and, given the specification, the model could potentially converge after one round of training. See the 112B rejections below for further explanation on the convergence limitation.
Claims 6 and 18 recite “…repeating the training a plurality of times until convergence” in various tense formats. This limitation leaves it unclear what is evaluated for convergence. Given the specification, the examiner can come to understand that training is repeated until the algorithm of the model is converged, however, the examiner suggests indicating this in the claims (e.g., “…repeating the training a plurality of times until the model has converged”). Additionally, the examiner notes that with no explicit definition of convergence in the specification, this limitation is open to a very broad interpretation. The examiner suggests explicitly defining convergence beyond one example embodiment that uses the loss function. For continued examination, the examiner can interpret convergence as any condition that terminates the training of a model.
Claims 7, 8, 10, 11, 19, 20, 22, 23 refer to the trained model as “the generative model, once trained,” “the trained Schrodinger-Bridge-based generative model,” “the Schrodinger-Bridge-based generative model,” and “the trained generative model.” The examiner asserts the importance of remaining consistent in the titling of elements that refer to the same thing. The examiner requires that for all references to the trained model only one of the above titles is used.
Claims 9 and 21 recite “the starting point is found by using the loss function that evaluates the data-log likelihood” and should read “the starting point is found by using the loss function which evaluates the data log-likelihood.” The recited corrections will make the limitation more grammatically correct and comprehensible in terms of its intended meaning given the specification.
Claims 11 and 23 recite “to predict the data-log likelihood” and should read “to predict the data log-likelihood” to fix the grammatical error
Claims 13 and 25 recite “using a reverse of the RH SDE solver…” and should read “using a reverse algorithm of the RH SDE solver…” in accordance with wording of the same limitation in claim 1, to remain consistent
Claim 25 recites “…to perform operations including configured to:…” should read “…to perform operations including:…” to fix the grammatical error
Claim 25 recites “…the computed initial values to a reversible Heun Stochastic Differential Equation (SDE) solver…” and should read “…the computed initial values to a Reversible-Heun (RH) Stochastic Differential Equation (SDE) solver…” to remain consistent with the other independent claims and fix the grammatical error. Additionally, claim 25 later recites “using a reverse of the reversible Heun…” and should read “using a reverse algorithm of the RH SDE solver…” in accordance with the other independent claims, provided that the issue above was correct as well.
Claims 2-12 and 14-24 further inherit the objections of the claims upon which they depend
Appropriate correction is required
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-25 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claims 1, 2, 13, 14, and 25 recite the limitation “clean example” or “clean image.” The examiner asserts that these limitations are indefinite as to what being “clean” means; the term is given no explicit definition and is referenced very few times in the specification. For continued examination, given the most reasonable understanding, the examiner comes to interpret “clean examples” or “clean images” to be types of data that are free of errors or extraneous data. Additions to the specification are required to more clearly define the scope of the term “clean” regardless of whether or not the interpretation above is accurate. Claims 2-12 and 14-24 further inherit this rejection from the claims upon which they depend.
Claims 1, 13, and 25 recite the limitation "feeding the initial values from the clean example…" in various tense formats. There is insufficient antecedent basis for this limitation in the claim because “the initial values of the clean example” are never introduced prior to this limitation. Given the specification, it cannot be assumed that the clean example inherently has initial values, so this limitation lacks antecedent basis; the examiner suggests including an introduction to the initial values of the clean example when the clean example is introduced during the sampling step (e.g., “…wherein the clean example comprises a set of initial values”). Claims 2-12 and 14-24 further inherit this rejection from the claims upon which they depend.
Claims 1, 13, and 25 further recite the limitation “feeding the initial values from the clean example and the computed initial values to…” in various tense formats. This limitation is considered indefinite because it can be misconstrued given that both sets of initial values are called “initial values”; One could reasonably misinterpret which initial values are which even with the “computed” modifier for the second set of initial values. The examiner suggests modifying the titles of these different elements with words like “first” and “second” in order to define the difference between them. One example alternate limitation could read: “feeding the first set of initial values from the clean example and the computed second set of initial values to a…” provided that the first and second set of initial values were introduced properly (i.e. as the “first” and “second” set) in the sampling and computing steps above this limitation. Claims 2-12 and 14-24 further inherit this rejection from the claims upon which they depend.
Claims 1, 13, and 25 further recite the limitation “…solving the stochastic adjoint SDE to…” in various tense formats. There is insufficient antecedent basis for this limitation in the claim because “the stochastic adjoint SDE” is never introduced prior to this limitation. Given the specification, the only indication as to what “the stochastic adjoint SDE” relates to, is that it is an equation of a method which uses an optimize-then-discretize approach to solve the SB-FBSDEs, however, there is no explicit definition of the equation or method or where its coming from, leaving the claims that recite this element lacking sufficient antecedent basis. The examiner requires that the actual equation and method are explicitly defined in the specification in order to provide sufficient antecedent basis for the term. For further examination, the examiner comes to interpret “the stochastic adjoint SDE” as being any equation solved in any “stochastic adjoint method.” Claims 2-12 and 14-24 further inherit this rejection from the claims upon which they depend.
Claims 1, 13, and 25 further recite the limitation “…to compute the gradient and update the weights of the generative model.” There is insufficient antecedent basis for this limitation in the claim because “the gradient” of the generative model is never introduced prior to this limitation and use of the word “the” makes it appear as it was; additionally, note that a gradient is not an inherit element of a generative model. The examiner comes to understand that the gradient is of the generative model but suggests rephrasing the limitation to “…to compute a gradient and update the weights of the generative model” so as it prevent potential misconstruction of the claim and address the issue with insufficient antecedent basis. Claims 2-12 and 14-24 further inherit this rejection from the claims upon which they depend.
Claims 3 and 15 recite the limitation “…utilizing the Nonlinear Feynman-Kac lemma to obtain…” There is insufficient antecedent basis for this limitation because “the Nonlinear Feynman-Kac lemma” is never introduced in the claims or specification prior to this limitation. The examiner comes to understand that the Feynman-Kac lemma is formula known to anyone of ordinary skill in the art, however, this does not mean that the applicant can reference it without explicitly defining what they understand to be the Feynman-Kac lemma and what parts of their invention are meant to represent the variables of the Feynman-Kac lemma. Further, in the specification, the lemma is only said to be “utilized” and “applied” but never is explicitly defined. The examiner requires that the applicant amends the specification to include their explicit definition of the Nonlinear Feynman-Kac lemma, including what variables of the Feynman-Kac lemma map to different parts of their method/invention.
Claims 4 and 16 recite the limitation “using a variant of the Stochastic Gradient Descent (SGD) deep learning optimizer” or “use a Stochastic Gradient Descent (SGD) deep learning optimizer.” There is insufficient antecedent basis for these limitations because “the Stochastic Gradient Descent (SGD) deep learning optimizer” is never introduced prior to these limitations in the claims and is never explicitly defined in the specification. The examiner comes to understand that concept of stochastic gradient descent in a foundational optimization technique known to anyone of ordinary skill in the art, however, this does not mean that the applicant can reference it as an optimizer without explicitly defining what they understand to be a “Stochastic Gradient Descent (SGD) deep learning optimizer.” The examiner requires that the applicant amends the specification to include their explicit definition of a “Stochastic Gradient Descent (SGD) deep learning optimizer,” including what variables of the SGD deep learning optimizer map to different parts of their method/invention. Additionally, note that the examiner suggests removing the wording “deep learning optimizer” from these limitations so as to prevent the applicant from further having to explain what a “deep learning optimizer” is (as its own particular element); an alternate representation of the limitation could read: “using a variant of Stochastic Gradient Descent (SGD) to update the weights…” provided that the applicant still defines what they understand to be “Stochastic Gradient Descent.” Claims 5 and 17 further inherit this rejection from the claims upon which they depend; Note that claims 5 and 17 also recite “deep learning optimizer” and should be amended if that phrasing is removed from claims 4 and 16.
Claim 4 further recites the limitation “…of the deep generative model using the gradients computed…” There is insufficient antecedent basis for this limitation because only a single gradient is computed in the limitations of claim 1. The examiner requires removing the pluralization of gradients in this limitation if it is to be understood that only a single gradient is computed, otherwise, the limitation of claim 1 must be modified to indicate computation of more than one gradient. Claim 5 further inherits this rejection from the claim upon which it depends.
Claims 5 and 17 recite the limitations “…wherein the SGD deep learning optimizer is the Adam optimizer” and “…wherein the SGD deep learning optimizer is an Adam optimizer.” There is insufficient antecedent basis for these limitations because “the Adam optimizer” is never introduced prior to these limitations in the claims and is never explicitly defined in the specification. The examiner comes to understand that the Adam optimizer is a known SGD optimizer to anyone of ordinary skill in the art, however, this does not mean that the applicant can reference it without explicitly defining what they understand to be an “Adam optimizer.” The examiner requires that the applicant amends the specification to include their explicit definition of “the Adam optimizer,” including, at least in the specification, how it is applied to their invention/method. Additionally, note that with the removal of “deep learning optimizer” from the limitations of claims 4 and 16 above, an alternate representation of the limitation could read: “…wherein the SGD is done using an Adam optimizer,” provided that the applicant explicitly defined what they understand to be an Adam optimizer in the specification.
Claims 8 and 20 recite the limitation “…to a data point with low data likelihood.” This limitation is indefinite because there is no definition for “data likelihood” in the specification. The examiner comes to understand that data likelihood may be a known concept to anyone of ordinary skill in the art, however, that does not mean the applicant can reference it without explicitly defining what they understand to be data likelihood. The examiner requires that the applicant amends the specification to include their explicit definition of data likelihood, including its application to these parts of their invention/method.
Claims 9, 11, 12, 21, 23, and 24 recite the limitation “…the data-log likelihood.” There is insufficient antecedent basis for this limitation in these claims because “the data-log likelihood” is never introduced prior to this limitation in the claims and is not explicitly defined in the specification. The examiner comes to understand that log-likelihood is a known concept to anyone of ordinary skill in the art, however, that does not mean the applicant can reference it without explicitly defining what they understand to be log-likelihood in the specification. The examiner requires that the applicant amends the specification to include their explicit definition of log-likelihood, including its specific application and mapping to the parts of their invention/method that use it. An alternate example limitation could read: “…is found by using the loss function that evaluates data log-likelihood,” (note removal of “the”) provided that the applicant explicitly defines what they understand to be log-likelihood in the specification. Further, note that in claims 12 and 24 the “data log-likelihood” is used in a manner in which to indicate that is was already calculated or determined, despite that fact that these claims depend on claims 8 and 20 which do not calculate or determine the data log-likelihood; note that claims 9, 11, 21, and 23 introduce the “data log-likelihood” in a manner that does not “use” it as if it has already been created. The examiner requires further rephrasing of these claims to indicate where this data log-likelihood was created, which may entail changing which claims these claims depend on.
Claims 12 and 24 further recite the limitation “…typicality test-based outlier detection scheme.” This limitation is indefinite because there is no definition of a “typicality test-based outlier detection scheme” in the specification. The examiner comes to understand that a typicality test-based outlier detection scheme may be a known concept to anyone of ordinary skill in the art, however, that does not mean the applicant can reference it without explicitly defining what they understand to be a typicality test-based outlier detection scheme. The examiner requires that the applicant amends the specification to include their explicit definition of a typicality test-based outlier detection scheme, including its application to these parts of their invention/method.
Claim 14 recites the limitation “…wherein the clean example…” There is insufficient antecedent basis for this limitation because only a “clean image” is introduced in claim 13, upon which claim 14 depends. The examiner notes that this issue should not persist provided that the applicant amends the claims to refer to only either a “clean image” OR a “clean example” throughout the disclosure, as suggested above.
Claim 16 recites the limitation “…to use a Stochastic Gradient Descent (SGD) deep learning optimizer to compute the gradient.” The examiner asserts that this claim is indefinite because “the gradient” is already computed in the limitations of claim 13, upon which claim 16 depends. Given the specification and other claims, the examiner believes that this is a typographical error and that the claim should be amended to recite: “…to use a Stochastic Gradient Descent (SGD) deep learning optimizer to update the weights of the generative model using the gradient computed using the reverse algorithm of the RH SDE solver” in accordance with the specification and the recitation of claim 4, of which claim 16 is an alternate embodiment. The examiner notes that if this is not a typographical error, the claim will be rejected under 112D for failing to incorporate all of the limitations of the claim upon which it depends. Claim 17 further inherits this rejection from the claim upon which it depends.
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 1-25 are rejected under 35 U.S.C. 101 because the claimed invention is directed towards an abstract idea without significantly more.
Step 1: Claims 1-12 are directed towards a method, Claims 13-24 are directed towards a machine, and Claim 25 is directed towards an article of manufacture. Therefore, Claims 1-25 are directed towards one of the 4 statutory categories: process, machine, article of manufacture, or composition of matter.
With respect to claim 1:
Step 2A Prong 1: The claim is directed to a judicial exception.
A method for training a Schrodinger-Bridge-based generative model, comprising… (Mental Process and Mathematical Concept: One could train a generative model to be a Schrodinger-Bridge-based generative model through repeated mental processing and application of mathematical concepts, mentally or using pen and paper)
sampling a clean example to be learned, the clean example being from an input set of training data to be used to train the generative model (Mental Process: One can sample a clean example to be learned from an input set of training data to be used to train the generative model, mentally or using pen and paper)
computing initial values using the sampled clean example… (Mental Process and Mathematical Concept: One could compute initial values using the sampled clean example by applying mathematical concepts, mentally or using pen and paper)
…to forward propagate using Schrodinger Bridge Forward-Backward Stochastic Differential Equations (SB-FBSDEs) computed for the generative modeling problem, producing predicted output values (Mental Process and Mathematical Concept: One could forward propagate using SB-FBSDEs that they computed for the generative modeling problem, to produce predicted output values, mentally or using pen and paper)
computing a loss function to compares true values and the predicted output values (Mental Process and Mathematical Concept: One could compute a loss function to compare true values and the predicted output values by applying mathematical concepts, mentally or using pen and paper)
…solving the stochastic adjoint SDE to compute the gradient and update the weights of the generative model (Mental Process and Mathematical Concept: One could solve the stochastic adjoint SDE to compute the gradient and update the weights of the generative model, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
computing…using… the generative model (High-level Machine Learning. Any model can be configured to be a Schrodinger-Bridge-based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
feeding the initial values from the clean example and the computed initial values to a Reversible-Heun (RH) Stochastic Differential Equation (SDE) solver (Amounts to necessary data output and gathering. Insignificant extra-solution activity, as discussed in MPEP § 2106.05(g))
using a reverse algorithm of the RH SDE solver… (High-level Machine Learning. Any algorithm can be configured to be a reverse algorithm of the RH SDE solver. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Re-evaluation of Insignificant Extra-Solution Activities:
feeding the initial values from the clean example and the computed initial values to a Reversible-Heun (RH) Stochastic Differential Equation (SDE) solver (“Receiving or transmitting data over a network” is a well-understood, routine, conventional activity when claimed in a merely generic manner (as it is in the present claim), as discussed in MPEP § 2106.05(d)(II))
Additional Elements:
computing…using… the generative model (High-level Machine Learning. Any model can be configured to be a Schrodinger-Bridge-based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
using a reverse algorithm of the RH SDE solver… (High-level Machine Learning. Any algorithm can be configured to be a reverse algorithm of the RH SDE solver. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 2:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
…wherein the clean example is a two-dimensional image (Mental Process: One could sample the clean example wherein the clean example is a two-dimensional image, mentally or using pen and paper. Adding information about what the abstract idea operates on does not integrate it into a practical application)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
With respect to claim 3:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
…further comprising utilizing the Nonlinear Feynman-Kac lemma to obtain the SB-FBSDEs corresponding to Schrodinger Bridge Partial Differential Equations (SB-PDEs) of the generative modeling problem (Mental Process and Mathematical Concept: One could use the Nonlinear Feynman-Kac lemma to obtain the SB-FBSDEs corresponding to Schrodinger Bridge Partial Differential Equations (SB-PDEs) of the generative modeling problem by applying mathematical concepts and using mental processes, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
With respect to claim 4:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
…further comprising using a variant of the Stochastic Gradient Descent (SGD) deep learning optimizer to update the weights of the deep generative model using the gradients (Mental Process and Mathematical Concept: One could use a variant of stochastic gradient descent to update the weights of the generative model using the gradients that were computed already by applying mathematical concepts, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
…computed using the reverse algorithm of the RH SDE solver (High-level Machine Learning. Any algorithm can be configured to be a reverse algorithm of the RH SDE solver that computes the gradients as above. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
…computed using the reverse algorithm of the RH SDE solver (High-level Machine Learning. Any algorithm can be configured to be a reverse algorithm of the RH SDE solver that computes the gradients as above. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 5:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 4 via dependency.
…wherein the SGD deep learning optimizer is the Adam optimizer (Mathematical Concept: Naming which SGD optimization technique is used merely tells one what mathematical concepts are applied)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 4… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 4… (See above)
With respect to claim 6:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
…further comprising repeating the training a plurality of times until convergence (Mental Process and Mathematical Concept: One could repeat the training claimed above a plurality of times until a convergence is reached by applying repeated mathematical concepts and mental processes, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
With respect to claim 7:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
…further comprising using the generative model, once trained, for generating a data sample (High-level Machine Learning. Merely using a trained generative model to perform a task for you does not integrate a practical application. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
…further comprising using the generative model, once trained, for generating a data sample (High-level Machine Learning. Merely using a trained generative model to perform a task for you does not integrate a practical application. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 8:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 1 via dependency.
…finding a starting point drawn from a prior distribution that leads to a data point with low data likelihood (Mental Process: One could find a starting point from a prior distribution that leads to a data point with low data likelihood, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 1… (See above)
…further comprising using the trained Schrodinger-Bridge-based generative model for outlier generation by… (High-level Machine Learning. Merely using a trained generative model to perform a task for you does not integrate a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 1… (See above)
…further comprising using the trained Schrodinger-Bridge-based generative model for outlier generation by… (High-level Machine Learning. Merely using a trained generative model to perform a task for you does not integrate a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 9:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 8 via dependency.
…wherein the starting point is found by using the loss function that evaluates the data-log likelihood (Mental Process and Mathematical Concept: One could find the starting point by using mathematical concepts to apply the loss function and evaluate the data log-likelihood, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 8… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 8… (See above)
With respect to claim 10:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 8 via dependency.
…wherein the starting point is found by using a learned data log-likelihood loss function (Mental Process and Mathematical Concept: One could find the starting point by using mathematical concepts to apply a learned data log-likelihood loss function, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 8… (See above)
…of the Schrodinger-Bridge-based generative model (High-level Machine Learning. Merely using a function provided by a generative model does not integrate the judicial exception into a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 8… (See above)
…of the Schrodinger-Bridge-based generative model (High-level Machine Learning. Merely using a function provided by a generative model does not integrate the judicial exception into a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 11:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 8 via dependency.
…predict the data-log likelihood which is subsequently used directly as a classifier of outlier status based on a predefined outlier threshold value (Mental Process and Mathematical Concept: One could predict the data log-likelihood and subsequently use it directly as a classifier of outlier status based on a predefined outlier threshold values by applying mathematical concepts, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 8… (See above)
…further comprising using the trained generative model to (High-level Machine Learning. Merely using a generative model to perform a task does not integrate the judicial exception into a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 8… (See above)
…further comprising using the trained generative model to (High-level Machine Learning. Merely using a generative model to perform a task does not integrate the judicial exception into a practical application. Any generative model can be trained to be a Schrodinger-Bridge based generative model. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claim 12:
Step 2A Prong 1: The claim is directed to a judicial exception, including those inherited from claim 8 via dependency.
…further comprising using the data log-likelihood in a typicality test-based outlier detection scheme applied to multi-data point queries (Mental Process and Mathematical Concept: One could use the data log-likelihood in a typicality test-based outlier detection scheme applied to multi-data point queries by applying mathematical concepts, mentally or using pen and paper)
Step 2A Prong 2: The judicial exceptions as a whole are not integrated into a practical
application.
Additional Elements:
The method of claim 8… (See above)
Step 2B: The claim does not include additional elements that amount to significantly more than the judicial exception.
Additional Elements:
The method of claim 8… (See above)
With respect to claim 13:
See the rejection for claim 1 above. Note that the primary difference between claim 1 and claim 13 is that claim 1 is directed towards a method whereas claim 13 is directed towards a system that performs that method. The minute differences in phrasing and wording used do not change the analysis of judicial exceptions provided above. The only additional element in claim 13 is analyzed as follows:
A system for training a Schrodinger-Bridge-based generative model, comprising:
one or more computing devices configured to… (Non-specific computing devices are merely generic computer components. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
With respect to claims 14-24:
See the rejections above for claims 2-12 respectively. Note that the primary difference between claims 2-12 and claims 14-24 is that claims 2-12 are directed towards a method whereas claims 14-24 are directed towards a system that performs that method. The minute differences in phrasing and wording used do not change the analysis of judicial exceptions provided above. Any of claims 14-24 that recite the additional limitation of “…wherein the one or more computing devices are further configured to…” are directed to the analysis of the computing devices provided in the rejection for claim 13 above. Additionally, note that claim 22 has an additional limitation in comparison to claim 10, which is directed to a judicial exception as follows:
…using a learned data log-likelihood loss function… directly as a classifier of outlier status based on a predefined outlier threshold value (Mental Process and Mathematical Concept: One could use a learned data log-likelihood loss function directly as a classifier of outlier status based on a predefined outlier threshold value by applying mathematical concepts and mental processes, mentally or using pen and paper)
With respect to claim 25:
See the rejection for claim 1 above. Note that the primary difference between claim 1 and claim 25 is that claim 1 is directed towards a method whereas claim 25 is directed towards an article of manufacture which holds instructions for that method. The minute differences in phrasing and wording used do not change the analysis of judicial exceptions provided above. The only additional element in claim 25 is analyzed as follows:
A non-transitory computer-readable medium comprising instructions for training a Schrodinger-Bridge-based generative model that, when executed by one or more computing devices, cause the one or more computing devices to perform operations including configured to… (Computer-readable mediums are generic computer components which all hold information including, potentially, instructions. Computing devices are also generic computing components as above. Adding generic computer components to perform the method is not sufficient. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f))
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-4 and 6-7 are rejected under 35 U.S.C. 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761)
Regarding claim 1, Chen et. al. teaches a method for training a Schrodinger-Bridge-based generative model (Abstract, “In this work, we present a novel computational framework for likelihood training of SB [Schrödinger-Bridge] models grounded on Forward-Backward Stochastic Differential Equations Theory… we show that the resulting training algorithm achieves comparable results on generating realistic images.” The examiner comes to understand that Chen is introducing a method for training a Schrodinger-Bridge based model which is used for generating images, i.e. the model is generative.), comprising:
sampling a clean… (Figure 1 and Figure 2, Note that in using a Schrodinger-Bridge based model, it is understood that a data to noise diffusion process is first learned by the model using “clean” examples. The examiner notes that with no explicit definition of “clean” it is understood to mean examples of data that are free from noise.) …example, to be learned… (3.3 Practical Implementation, Algorithm 2 or 3, “Sample Xt ∈ [0, T] from (13a) where x0 ∼ pdata.” The examiner comes to understand that at this step in the training process of Algorithms 2 and 3, an example, X0, is sampled from the input set of data, pdata which is set to be used in the forward propagation of 13a to get Xt. See figure 2, showing that X0 is a “clean” example as the first training process (forward propagation) of SB goes from data to noise.), …the clean example being from an input set of training data to be used to train the generative model (3.3 Practical Implementation, Algorithm 1, “Input: boundary distributions pdata and pprior.” The examiner comes to understand that in the training process two data distributions are input to train the model and, as above, it is understood that the clean example, X0, comes from pdata, i.e. the clean example comes from an input set of training data.)
computing initial values… (3.1 Forward-Backward SDEs (FBSDEs) Representation For SB, Theorem 3, Equations 13a-13c. The examiner comes to understand that, as above, when Xt is sampled in Algorithms 2 and 3, this system of equations, 13a-13c, is solved, computing initial values of the model, using the clean example X0. These computed initial values consist of Y0 and Ŷ0.) using the sampled clean example and the generative model (Theorem 3, Note that these equations are made for the SB model)
…using Schrodinger Bridge Forward-Backward Stochastic Differential Equations (SB-FBSDEs) computed for the generative modeling problem… (3.1 Forward-Backward SDEs (FBSDEs) Representation For SB and Theorem 3, “Below we derive a similar FBSDEs representation for SB…, [Equations 13a-13c], …The FBSDEs for SB (13) share a similar forward-backward structure as in (9), where (13a) and (13b,13c) respectively represent the forward and backward SDEs… In other words, these FBSDEs provide… Zt and Ẑt can be understood as the forward/backward policies, in a similar spirit of policy-based methods (Pereira et al., 2020; Schulman et al., 2015), that guide the SDE processes of SB, they sufficiently characterize the SB model.” The examiner notes that the equations, 13a-13c, are understood to be Schrodinger Bridge Forward-Backward Stochastic Differential Equations (SB-FBSESs) computed for the generative SB model and modeling problem.)
computing a loss function to compares true values and the predicted output values (3.3 Practical Implementation, Algorithms 2 and 3, “Compute LSB(x0; θ, φ) with (16)” and Equations 16, 18, and 19. The examiner comes to understand that the training algorithms use computed loss functions, Equation 16, 18 and 19, which evaluats log-likelihood of data to compare true values and predicted output values. The examiner notes that log-likelihood measures how well a model explains observed data which requires comparing true values and predicted values.)
and…
…update the weights of the generative model (3.3 Practical Implementation, Algorithms 2 and 3, “Update φ with gradient…,” “Update θ with gradient…,” and “Update (θ, φ) with…” The examiner comes to understand that in these steps of the algorithms, the weights of the model φ and θ, are updated with a computed gradient of the model)
Chen et. al. does not distinctly disclose:
feeding the initial values from the clean example and the computed initial values to a Reversible-Heun (RH) Stochastic Differential Equation (SDE) solver to forward propagate… OR using a reverse algorithm of the RH SDE solver, solving the stochastic adjoint SDE to compute the gradient…
However, Kidger et. al. teaches those limitations:
feeding the initial values from the clean example and the computed initial values to a Reversible-Heun (RH) Stochastic Differential Equation (SDE) solver to forward propagate… (3 Reversible Heun Method, Algorithm 1: Forward Pass, “We introduce a new SDE solver, which we refer to as the reversible Heun method. Its key property is algebraic reversibility; moreover, to the best of our knowledge it is the first SDE solver to exhibit this property. To fix notation, we consider solving the Stratonovich SDE… Solver We begin by selecting a step size ∆t, and initializing t0 = 0, z0 = ẑ0 = Z0, µ0 = µ(0, Z0) and σ0 = σ(0, Z0)… We then iterate Algorithm 1.” The examiner comes to understand that this Reversible Heun solver can be applied to the problem determined in Chen to determine Xt instead of using the forward propagation techniques used by Chen. In this application, one can feed the initial values, X0, and the computed initial values Y0 and Ŷ0 along with the computed forward and backward policies Zt and Ẑt from Equations 14 in Chen to forward propagate Xt, Yt, and Ŷt; In this application, the input values of Algorithm 1, tn, zn, ẑn, µn, and σn are replaced by the corresponding values listed above. The examiner notes that the Reversible Heun solver can be used to solve Stratonovich SDEs, so to solve the set of SB-FBSDEs, determined above, one can convert these equations from Itô to Stratonovich type because Itô is less efficient as noted in Appendix C.)
…producing predicted output values (3 Reversible Heun Method and Algorithm 1: Forward Pass, “Suppose T = N∆t so that zN, ẑN, µN, σN are the final output. Then zN ≈ZT is returned.” The examiner comes to understand that when the change in time in Algorithm 1, for the forward pass, is set to N∆t, the terminal output values are determined, which are understood to be the predicted output values.)
and…
using a reverse algorithm of the RH SDE solver… (3 Reversible Heun Method, Algorithm 2, “whilst zN, ẑN, µN, σN are all retained for the backward pass.” The examiner comes to understand that Algorithm 2 of the Reversible Heun method is the backward pass, or reverse algorithm of the solver which is used with the forward propagated values from Algorithm 1, a.k.a. the forward pass), …solving the stochastic adjoint SDE (3 Reversible Heun Method, Algebraic Reversibility, “The key advantage of the reversible Heun method, and the motivating reason for its use alongside continuous-time adjoint methods, is that it is algebraically reversible…. will mean that it is possible to backpropagate through the SDE solve, such that the gradients obtained via the continuous adjoint method...” The examiner notes that given the specification, the only defining quality of the “stochastic adjoin SDE”, is that it is solved using a stochastic adjoint method which is an optimize-then-discretize approach used to solve the SB-FBSDEs. The examiner comes to find this definition to be in accordance with Kidger’s definition of the same: “Here we use the continuous adjoint method. Also known as simply ‘the adjoint method’, or ‘optimize then-discretize.’” This means that the Reversible Heun Solver solves the “stochastic adjoin SDE” using the continuous adjoint method. Additionally, note that the examiner comes to understand that the word “stochastic” added to this adjoint SDE and method does not change what is already known about the problem.) …to compute the gradient… (3 Reversible Heun Method, Algorithm 2, Output, note that the output contains the gradients computed using the reverse algorithm of the RH solver by solving via the adjoint method the SB-FBSDEs of the model. Further, note that when combined with Chen et. al., these computed gradients can be used to update those of the generative model as in Algorithms 2 and 3 of Chen et. al.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) with the technique for using a Reversible-Heun SDE solver in Kidger et. al. in order to reduce the computation complexity of solving the SB-FBSDEs and decrease the time required to train the model. (Kidger et. al., 3 Reversible Heun Method, Computation Efficiency, “A further advantage of the reversible Heun method is computational efficiency. The method requires only a single function evaluation (of both the drift and diffusion) per step. This is in contrast to other Stratonovich solvers (such as the midpoint method or regular Heun’s method), which require two function evaluations per step,” and see Table 1 which corresponds to 3.1 Experiments that compare the Reversible Heun Solver to other known methods; note that the training time for the Reversible Heun method is significantly faster than other known methods.)
Regarding claim 2, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 1 as cited above and Chen et. al. further teaches the limitation:
…wherein the clean example is a two-dimensional image (4 Experiments, Setups, “We testify SB-FBSDE on two toy datasets and three image datasets, i.e. MNIST, CelebA, 6 and CIFAR10,” see Figure 4 which has examples of the data in datasets CelebA and CIFAR10. Note that the examiner comes to interpret these images in the datasets as 2-dimensional images. Further, note that these datasets are understood to partially or fully make up pdata from which the clean example is sampled, as above.)
Regarding claim 3, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 1 as cited above and Chen et. al. further teaches the limitation:
…further comprising utilizing the Nonlinear Feynman-Kac lemma to obtain the SB-FBSDEs corresponding to Schrodinger Bridge Partial Differential Equations (SB-PDEs) of the generative modeling problem (3.1 Forward-Backward SDEs (FBSDEs) Representation For SB, Lemma 2: The Nonlinear Feynman-Kac lemma and Theorem 3, “Hence, one can adopt Lemma 2 to solve the underlying FBSDEs, rather than the original PDE optimality, for the optimal control… Below we derive a similar FBSDEs representation for SB… Consider the following set of coupled SDEs… where f and g satisfy the same regularity conditions in Lemma 2… then the nonlinear Feynman-Kac relations between the FBSDEs (13) and PDEs (6) are given by…,” see equation set 14. Further, see Appendix B which shows the proof for creating Equations 13a-13c (the SB-FBSDEs) wherein the equations are created “by rewriting (33) and (36) with the nonlinear Feynman-Kac” lemma. The examiner notes that these SB-FBSDEs are created to characterize the SB model and thereby correspond to the modeling problem and the SB-PDEs.)
Regarding claim 4, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 1 as cited above and Chen et. al. further teaches the limitation:
…further comprising… update the weights of the deep generative model using the gradients computed (3.3 Practical Implementation, Algorithms 3, “Update φ with gradient ∇φ LSB(x0; φ)” and “Update θ with gradient ∇θ LSB(xT; θ).” The examiner comes to understand that, as above, the weights of the model are updated during these steps of training using the computed gradients which, as determined above, can be replaced by those computed using the RH SDE solver instead of the ones listed)
Chen et. al. does not distinctly disclose:
…further comprising using a variant of the Stochastic Gradient Descent (SGD) deep learning optimizer to update the weights… using the gradients computed using the reverse algorithm of the RH SDE solver
However, Kidger et. al. teaches these limitations:
…further comprising using a variant of the Stochastic Gradient Descent (SGD) deep learning optimizer to update the weights… (3.1 Experiments, Versus Midpoint, “We begin by comparing the reversible Heun method with the midpoint method, which also converges to the Stratonovich solution. We train an SDE-GAN on a dataset of weight trajectories evolving under stochastic gradient descent, and train a Latent SDE on a dataset of air quality over Beijing.” The examiner comes to understand this to mean that the Reversible Heun method uses stochastic gradient descent to update the weights of the model. Note that without an explicit definition for “optimizer” the examiner comes to interpret it as comprising any element or step that performs stochastic gradients descent, of which, the citation above provides.) …using the gradients computed using the reverse algorithm of the RH SDE solver (3 Reversible Heun Method, Algorithm 2, Output, The outputs of the model are understood to be the gradients computed using the Reversible Heun method. As noted above, one could use the gradients computed by the Reversible Heun method instead of those computed in Chen et. al. to update the weights of the generative model outlined by Chen et. al. in algorithms 2 and 3; Further, in combination with the citation above, those weights can be updated using stochastic gradient descent as in Kidger et. al.)
See the rationale for combining Chen et. al. with Kidger et. al. in the rejection for claim 1 above. Note that Reversible Heun (RH) method taught in Kidger et. al. comprises using stochastic gradient descent to update the weights of the model using the gradients computed by the RH solver, therefore, the rationale to combine remains the same.
Regarding claim 6, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 1 as cited above and Chen et al. further teaches the limitation:
…further comprising repeating the training a plurality of times until convergence (3.3 Practical Implementation, Algorithm 1: Likelihood Training of SB-FBSDE, “repeat: if memory resource is affordable, then run Algorithm 2, else run Algorithm 3, end if, until converges.” The examiner comes to understand that the training is repeated via Algorithms 2 and 3 a plurality of times until the model converges.)
Regarding claim 7, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 1 as cited above and Chen et. al. further teaches the limitation:
…further comprising using the generative model, once trained, for generating a data sample (4 Experiments, Image Datasets, “Next, we validate our alternate training (i.e. Alg 3) on high-dimensional image generation. The images generated for MNIST, CelebA, and CIFAR10 are presented in Fig. 4, which clearly suggest that our SB-FBSDE is able to synthesize high-fidelity images.” From this, the examiner comes to understand that the trained model is used to generate images which are data samples.)
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761), further in view of Kingma et. al. (Kingma, Diederik P., and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).)
Regarding claim 5, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of the method of claim 4 as cited above and Kingma et. al. further teaches the limitation:
…wherein the SGD deep learning optimizer is the Adam optimizer (1 Introduction and Algorithm 1: Adam, “many objective functions are composed of a sum of subfunctions evaluated at different subsamples of data; in this case optimization can be made more efficient by taking gradient steps w.r.t. individual subfunctions, i.e. stochastic gradient descent (SGD) or ascent… We propose Adam, a method for efficient stochastic optimization that only requires first-order gradients with little memory requirement.” The examiner comes to understand that this Adam optimization, which relies on a stochastic gradient descent approach, is “the Adam optimizer” used in the claimed invention.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) as modified with the technique for using a Reversible-Heun SDE solver in Kidger et. al. with the Adam optimizer introduced in Kingma et. al. in order to efficiently update the weights of the model without using a lot of memory and address issues with sparse gradients and non-stationary objectives. (Kingma, 8. Conclusion, “The method combines the advantages of two recently popular optimization methods: the ability of AdaGrad to deal with sparse gradients, and the ability of RMSProp to deal with non-stationary objectives. The method is straightforward to implement and requires little memory.”)
Claims 8-11 are rejected under 35 U.S.C 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761) and further in view of Kumar et. al. (Kumar, Nishant, et al. "Normalizing flow based feature synthesis for outlier-aware object detection." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.)
Regarding claim 8, Chen et. al. as modified by Kidger et. al. teaches all of the limitations of claim 1 as cited above and Chen et. al. further teaches the limitation:
…further comprising using the trained Schrodinger-Bridge based generative model for… generation… (3.3 Practical Implementation, Generative Process, “While the generative processes for SB can be performed as simply as propagating (7b) given the trained policy Z(·,·;φ), it has been constantly observed that adopting Langevin sampling to the generative process greatly improves performance…This results in the following predictor-corrector sampling procedure (see Alg. 4 in Appendix D for more details)…,” See equations 20 and 21. The examiner comes to understand that Langevin Sampling is a generative sampling technique.) …by finding a starting point drawn from a prior distribution… (Appendix D, Algorithm 4: Generative Process for SB-FBSDE, “Sample XT ∼ pprior.” The examiner comes to understand that the Langevin generative sampling technique uses samples from pprior which is a prior distribution from which a starting point is determined, this starting point is XT.)
Chen et. al. does not distinctly disclose:
…outlier generation by finding a starting point… that leads to a data point with low data likelihood
However, Kumar et. al. teaches this limitation:
…outlier generation (3.2.4 Projection Sampling Using Langevin Dynamics, “To accomplish this, we perform projection sampling based on Stochastic Gradient Langevin Dynamics (SGLD) to sample outliers from our flow model. Under the approach, we generate s samples as outliers os…” The examiner comes to understand the approach taught to be equivalent to outlier generation.) …by finding a starting point… that leads to a data point with low data likelihood (3.2.4 Projection Sampling Using Langevin Dynamics, “Under the approach, we generate s samples as outliers os directly and require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score δ of the inlier, which is least likely to be obtained from the inlier data distribution pγ(l). Next, we propagate the average log-likelihood score of the synthetic outliers logpγ(os) in the inverse direction of the flow without updating its parameters and obtain the gradients ∂logpγ(os)/∂os. Subsequently, the outliers os are updated as: [Equation 3] where T is the step size of the gradient descent. We end this iterative process when the average log-likelihood of the updated os is equal to or lower than δ. Hence, the updated outliers os are precisely projected to the near decision boundary of the inlier manifold where the least likelihood inlier is located.” The examiner comes to understand that this process can be applied to the Langevin generative sampling technique in Chen et. al. in order to generate outliers that lead to data points with low data likelihoods. In this combination, the starting point XT from the prior distribution pprior can be the s sample for the initial outliers os. As noted in the citation above, this outlier os (a.k.a. the starting point) is iteratively updated with recomputed gradients in order to find an outlier os with low data likelihood.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) as modified with the technique for using a Reversible-Heun SDE solver in Kidger et. al. with the Stochastic Gradient Langevin Dynamics sampling strategy proposed in Kumar et. al. in order to generate outliers near a decision boundary that corresponds to low data likelihood. (Kumar et. al., 3.2.4 Projection Sampling Using Langevin Dynamics, “Hence, the updated outliers os are precisely projected to the near decision boundary of the inlier manifold where the least likelihood inlier is located.”)
Regarding claim 9, Chen et. al. as modified by Kidger et. al. and Kumar et. al. teaches all of the limitations of claim 8 as cited above and Chen et. al. further teaches the limitation:
…the loss function that evaluates the data-log likelihood (3.2 Log-likelihood Computation of SB, Theorem 4, “Given the solution satisfying the FBSDE system in (13), the log-likelihood of the SB model (Zt, Ẑt), at a data point x0, can be expressed as…,” see Equation 16. Additionally, see equations 18 and 19. The examiner comes to understand that Equations 16, 18, and 19 are loss functions, used in training algorithms 2 and 3, which evaluate the data log-likelihood of a data point x0 or xT of the SB model. Note that these are understood to be the same loss functions previously introduced in claim 1 and cover what is deemed as “the loss function.”)
Chen et. al. does not distinctly disclose:
…wherein the starting point is found by using the… function that evaluates the data-log likelihood
However, Kumar et. al. teaches this limitation:
…wherein the starting point is found by using the… function that evaluates the data-log likelihood (3.2.4 Projection Sampling Using Langevin Dynamics, as in the interpretation for claim 8 above the starting point XT from Chen et. al. can be used as the s sample for the initial outlier os . This starting point is found by iteratively updating its values using the “average log-likelihood score of the synthetic outliers logpγ(os)” to obtain and use gradients, until a threshold log-likelihood is reached. The examiner comes to understand that in combination with Chen et. al. one could simply replace the log-likelihood function used to propagate the outliers in Kumar et. al. with “the” log-likelihood loss function computed in Chen et. al. as interpreted above.)
See the rationale for combining Chen et. al. as modified by Kidger et. al. with Kumar et. al. in the rejection for claim 8 above. Note that the Stochastic Gradient Langevin Dynamics sampling strategy used in Kumar et. al. includes using a log-likelihood function to find the starting point for outlier generation and, therefore, the rationale is the same.
Regarding claim 10, Chen et. al. as modified by Kidger et. al. and Kumar et. al. teaches all of the limitations of claim 8 as cited above and Chen et. al. further teaches the limitation:
…a learned data log-likelihood loss function of the Schrodinger-Bridge-based generative model (3.2 Log-likelihood Computation of SB, Theorem 4, “Given the solution satisfying the FBSDE system in (13), the log-likelihood of the SB model (Zt, Ẑt), at a data point x0, can be expressed as…,” see Equation 16. Additionally, see equations 18 and 19. The examiner comes to understand that Equations 16, 18, and 19 are loss functions, used in training algorithms 2 and 3, which evaluate the data log-likelihood of a data point x0 or xT of the SB model. Note that the equations rely on the learned forward and backward policies of the generative model and are thereby also “learned” and “of” the Schrodinger-Bridge-based generative model. Any of Equations 16, 18, and 19 can be interpreted as “a” learned data log-likelihood loss function of the model.)
Chen et. al. does not distinctly disclose:
…wherein the starting point is found by using a… data log-likelihood… function…
However, Kumar et. al. discloses this limitation:
…wherein the starting point is found by using a… data log-likelihood… function… (3.2.4 Projection Sampling Using Langevin Dynamics, as in the interpretation for claim 8 above the starting point XT from Chen et. al. can be used as the s sample for the initial outlier os . This starting point is found by iteratively updating its values using the “average log-likelihood score of the synthetic outliers logpγ(os)” to obtain and use gradients, until a threshold log-likelihood is reached. The examiner comes to understand that in combination with Chen et. al. one could simply replace the log-likelihood function used to propagate the outliers in Kumar et. al. with a learned data log-likelihood loss function in Chen et. al. as interpreted above.)
See the rationale for combining Chen et. al. as modified by Kidger et. al. with Kumar et. al. in the rejection for claim 8 above. Note that the Stochastic Gradient Langevin Dynamics sampling strategy used in Kumar et. al. includes using a log-likelihood function to find the starting point for outlier generation and, therefore, the rationale is the same.
Regarding claim 11, Chen et. al. as modified by Kidger et. al. and Kumar et. al. teaches all of the limitations of claim 8 as cited above and Chen et. al. further teaches the limitation:
…further comprising using the trained generative model to predict the data-log likelihood (3.2 Log-likelihood Computation of SB, Theorem 4, The examiner comes to understand that this theorem provides Equations 15 and 16 which use the learned forward and backward policies, ZT and ẐT, of the trained model to predict the data log-likelihood of a data point, x0.)
Chen et. al. does not distinctly disclose:
…the data-log likelihood which is subsequently used directly as a classifier of outlier status based on a predefined outlier threshold value
However, Kumar et. al. discloses this limitation:
…the data-log likelihood which is subsequently used directly as a classifier of outlier status based on a predefined outlier threshold value (3.2.4 Projection Sampling Using Langevin Dynamics, “Under the approach, we… require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score δ of the inlier, which is least likely to be obtained from the inlier data distribution pγ(l)… We end this iterative process when the average log-likelihood of the updated os is equal to or lower than δ. Hence, the updated outliers os are precisely projected to the near decision boundary of the inlier manifold where the least likelihood inlier is located.” The examiner comes to understand that in this citation the average log-likelihood of the outliers, os, is used directly as a classifier of outlier status based on the predefined outlier threshold δ; The average log-likelihood of the updated outliers is compared to the threshold value, δ, to determine the status of the outliers as either close enough or not close enough to the desired low data likelihood. Given no explicit definition in the specification, the status of the outliers can be interpreted with the above as being either close enough or not close enough to a desired low data likelihood. The examiner notes that the defined log-likelihood threshold value is understood to be a predefined outlier threshold value as it is determined before the samples are iteratively updated. In combination with Chen et. al., the log-likelihood function used as the classifier in Kumar et. al. could be replaced by the log-likelihood function determined in Chen et. al., as interpreted above, in order apply the classifier process to the trained generative model and its corresponding data log-likelihood function.)
See the rationale for combining Chen et. al. as modified by Kidger et. al. with Kumar et. al. in the rejection for claim 8 above. Note that the Stochastic Gradient Langevin Dynamics sampling strategy used in Kumar et. al. includes comparison of the computed log-likelihoods to a threshold value and, therefore, the rationale is the same.
Claim 12 is rejected under 35 U.S.C 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761), further in view of Kumar et. al. (Kumar, Nishant, et al. "Normalizing flow based feature synthesis for outlier-aware object detection." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.), further in view of Nalisnick et. al. (Nalisnick, Eric, et al. "Detecting out-of-distribution inputs to deep generative models using typicality." arXiv preprint arXiv:1906.02994 (2019).)
Regarding claim 12, Chen et. al. as modified by Kidger et. al. and Kumar et. al. teaches all of the limitations of claim 8 as cited above and Chen et. al. further teaches the limitation:
…the data log-likelihood (3.2 Log-likelihood Computer of SB, Theorem 4, As interpreted above, the data log-likelihood of the trained model can be computed for a data point, x0, using Equation 16. Note that the examiner comes to interpret use of the term “the” as reference to the data log-likelihood of the model.)
Chen et. al. does not distinctly disclose:
…using… data log-likelihood in a typicality test-based outlier detection scheme applied to multi-data point queries.
However, Nalisnick et. al. teaches this limitation:
…further comprising using… data log-likelihood in a typicality test-based outlier detection scheme applied to multi-data point queries (Appendix B, Algorithm 1: Bootstrap Test of Typicality. The examiner comes to understand that this algorithm provides a scheme for using a log-likelihood function to determine out-of-distribution or outlier inputs to a model. The algorithm is typicality test based and applied to an M-sized batch of possibly OOD inputs X̃ which the examiner understands to be a multi-data point query. See line 2 of the algorithm which computes the data log-likelihood of the Xn inputs. Given no explicit definition in the specification, this typicality test OOD scheme can be interpreted as a typicality test-based outlier detection scheme. In combination with Chen et. al., one could simply replace the data log-likelihood function used in Nalisnick et. al. with the one defined in Chen et. al., as interpreted above, in order to apply this typicality test to the trained generative model.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) as modified with the technique for using a Reversible-Heun SDE solver in Kidger et. al. and further modified by the Stochastic Gradient Langevin Dynamics sampling strategy proposed in Kumar et. al. with the typicality test outlier detection scheme in Nalisnick et. al. in order to identify sets of outlier data points with high accuracy in a deep generative model. (Nalisnick, 6. Discussion and Conclusions, “In the experiments we showed that the proposed test is especially well-suited to DGMs, identifying the OOD set for SVHN vs CIFAR-10 vs ImageNet (Nalisnick et al., 2019) with high accuracy (while maintaining ≤ 1% type-I error).”)
Claims 13-16, 18-19, and 25 are rejected under 35 U.S.C. 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761) and further in view of Yoon et. al. (US 20230394319 A1).
Regarding claim 13, see the rejection for claim 1 above. Note that the primary difference between claim 1 and claim 13 is that claim 1 is directed towards a method whereas claim 13 is directed towards a system that performs that method. The minute differences in phrasing and wording used do not change interpretation of the claim with respect to the references used in the rejection for claim 1. Further, note that Yoon et. al. teaches the limitation including: A system for training a generative model… comprising: one or more computing devices configured to… (Paragraph [0013], “In accordance with another aspect of the present invention, there is provided an apparatus for training an… generative model, the apparatus including a processor, and a memory operably connected to the processor to store at least one piece of code executed by the processor.” The examiner comes to understand that a generative model is trained using a system comprising one or more computing devices (in one embodiment). In application to the present invention and in combination with the rejection of claim 1, one could train a Schrodinger-Bridge based generative model using a system with one or more computing devices.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) as modified with the technique for using a Reversible-Heun SDE solver in Kidger et. al. with the system for training a generative model in Yoon et. al. in order to perform the model training in an efficient and accurate environment, such as the one provided by a system with one or more computing devices.
Regarding claim 14, see the rejection for claim 2 above. Note that the primary difference between claim 2 and claim 14 is that claim 2 is directed towards a method whereas claim 14 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al.
Regarding claim 15, see the rejection for claim 3 above. Note that the primary difference between claim 3 and claim 15 is that claim 3 is directed towards a method whereas claim 15 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 15.
Regarding claim 16, see the rejection for claim 4 above. Note that the primary difference between claim 4 and claim 16 is that claim 4 is directed towards a method whereas claim 16 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 16. Additionally, note that as in the 112B rejection above, the examiner came to the belief that claim 16 had typographical errors. In examination of claim 16, the examiner comes to interpret its limitations to be in accordance with those of claim 4.
Regarding claim 18, see the rejection for claim 6 above. Note that the primary difference between claim 6 and claim 18 is that claim 6 is directed towards a method whereas claim 18 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 18.
Regarding claim 19, see the rejection for claim 7 above. Note that the primary difference between claim 7 and claim 19 is that claim 7 is directed towards a method whereas claim 19 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 19.
Regarding claim 25, see the rejection for claim 1 above. Note that the primary difference between claim 1 and claim 25 is that claim 1 is directed towards a method whereas claim 25 is directed towards an article of manufacture holding instructions for that method. The minute differences in phrasing and wording used do not change interpretation of the claim with respect to the references used in the rejection for claim 1. Further, note that Yoon et. al. teaches the limitation including: A non-transitory computer-readable medium comprising instructions for training a… generative model that, when executed by one or more computing devices, cause the one or more computing devices to perform operations including configured to:… (Paragraph [008], “it is an object of the present invention to provide a method and apparatus for training a generative model” and Paragraph [0049], “The memory 120 may store software and may include a volatile or nonvolatile recording medium. In addition, the memory 120 is connected to one or more processors 140 through an electrical or internal communication interface, and when executed by the processor 140, and may store code that causes the processor 140 or the learning processor 150 to train a deep learning-based learning model.” The examiner comes to understand that a generative model is trained using a computer-readable medium comprising instructions which are executed by a processor (in one embodiment). In application to the present invention and in combination with the rejection of claim 1, one could train a Schrodinger-Bridge based generative model using a computer-readable medium storing instructions for training the model which are executed by a processor.)
Before the effective filing date of the claimed invention it would have been obvious to one of ordinary skill in the art to combine the method for training a Schrodinger-Bridge based generative model in Chen et. al. (Including: sampling, computing initial values, computing SB-FBSDEs for the model, computing a loss function, and updating the weights of the model) as modified with the technique for using a Reversible-Heun SDE solver in Kidger et. al. with the system for training a generative model in Yoon et. al. in order to perform the model training in an efficient and accurate environment, such as the one provided by a memory holding instructions which are executed by a processor.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761), further in view of Kingma et. al. (Kingma, Diederik P., and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014), further in view of Yoon et. al. (US 20230394319 A1).
Regarding claim 17, see the rejection for claim 5 above. Note that the primary difference between claim 5 and claim 17 is that claim 5 is directed towards a method whereas claim 17 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al.
Claims 20-23 are rejected under 35 U.S.C 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761), further in view of Kumar et. al. (Kumar, Nishant, et al. "Normalizing flow based feature synthesis for outlier-aware object detection." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.), further in view of Yoon et. al. (US 20230394319 A1).
Regarding claim 20, see the rejection for claim 8 above. Note that the primary difference between claim 8 and claim 20 is that claim 8 is directed towards a method whereas claim 20 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 20.
Regarding claim 21, see the rejection for claim 9 above. Note that the primary difference between claim 9 and claim 21 is that claim 9 is directed towards a method whereas claim 21 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al.
Regarding claim 22, see the rejections for claims 10 and 11 above. Note that the primary difference between claims 10 and 11 and claim 22 is that claims 10 and 11 are directed towards a method whereas claim 22 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that claim 22 defers from claim 10 by adding the additional limitation: “…directly as a classifier of outlier status based on a predefined outlier threshold value.” This additional limitation is addressed in the rejection for claim 11 via Kumar et. al. and the rejection can further be applied to its recitation in this claim without change in interpretation from the rejection in claim 11.
Regarding claim 23, see the rejection for claim 11 above. Note that the primary difference between claim 11 and claim 23 is that claim 11 is directed towards a method whereas claim 23 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 23.
Claim 24 is rejected under 35 U.S.C 103 as being unpatentable over Chen et. al. (Chen, Tianrong, Guan-Horng Liu, and Evangelos A. Theodorou. "Likelihood training of Schrödinger Bridge using forward-backward sdes theory." arXiv preprint arXiv:2110.11291 (2021)) in view of Kidger et. al. (Kidger, Patrick, et al. "Efficient and accurate gradients for neural sdes." Advances in neural information processing systems 34 (2021): 18747-18761), further in view of Kumar et. al. (Kumar, Nishant, et al. "Normalizing flow based feature synthesis for outlier-aware object detection." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.), further in view of Nalisnick et. al. (Nalisnick, Eric, et al. "Detecting out-of-distribution inputs to deep generative models using typicality." arXiv preprint arXiv:1906.02994 (2019)), further in view of Yoon et. al. (US 20230394319 A1).
Regarding claim 24, see the rejection for claim 12 above. Note that the primary difference between claim 12 and claim 24 is that claim 12 is directed towards a method whereas claim 24 is directed towards a system that performs that method, introduced in claim 13; See the rejection for claim 13 above which addresses the alternate embodiment as a system, including the rationale for incorporating Yoon et. al. Additionally, note that Yoon et. al., as used in the rejection for claim 13, covers the additional use of the computing devices configured to perform the limitations of claim 24.
Citation of Pertinent Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's
Disclosure. Pereira et. al. teaches deep neural network architectures for stochastic control in application to solving non-linear stochastic systems and decision-making problems for robotics and autonomy. These architectures reformulate the stochastic control policy in terms of FBSDEs. Additionally, their methods compute loss functions to compare true values to predicted values and make use of the Adam optimizer to update the weights of the model. Wang et. al. discloses a method for learning a Schrodinger-Bridge based generative model via entropy interpolation. Their method uses a time-varying drift term in solving their SDEs and the experimental phase of their teachings is very similar to that of the proposed invention. Bortoli et. al. teaches a diffusion-based Schrodinger-Bridge generative model that makes use of reverse time SDEs. Liu et. al. discloses an application of Schrodinger-Bridge based deep generative modeling to solve mean-field games with distributional boundary constraints. The disclosure includes computations of FBSDEs using the non-linear Feynman-Kac lemma as well as an algorithm for training which is very similar to the one taught in Chen et. al. Chen et. al. (2) teaches an algorithm for likelihood training of a Schrodinger-Bridge based generative model which also coincides with the algorithms used in Chen et. al. (1). Du et. al. discloses use of a diffusion model to generate outlier data which is in the low-likelihood region of the latent feature space. The original disclosure is framed for feature sets but can be generalized in application to image classification. Du et. al. (2) teaches a novel framework for out-of-distribution detection by synthesizing outliers from a low-likelihood data region of a feature space. This framework can additionally be formulated as a binary classification problem in terms of classifying the status of objects as being in-distribution or out-of-distribution. The disclosure also makes use of a thresholding comparison step in order to compare the synthesized outlier data to a certain metric. Song et. al. provides a methodology for more efficient training of score based generative models and incorporates annealed Langevin dynamics which could be applied to generation of samples from complex distributions, like outliers.
Conclusion
The prior art made of record and not relied upon is considered pertinent to Applicant's
disclosure. The applicant is required under 37 C.F.R. § 1.111(c) to consider these references fully when responding to this action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Zane A Rawlings whose telephone number is (571)270-3372. The examiner can normally be reached M-F, 8am to 5pm ET.
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, Alexey Shmatov can be reached at (571) 270-3428. 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.
/Z.A.R./Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights