Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 02/24/2026 has been entered.
Response to Remarks
Claim Rejections – 35 U.S.C. 101
Applicant’s arguments have been fully considered but they are not persuasive.
Applicant argues (pg. 9): “Examiner’s analysis under Step 2B improperly truncates the claimed invention by addressing only an arbitrary subset of the recited limitations. Specifically, the Examiner’s analysis considers only step (b) …, only a portion of step (c) …, and step (g) …. The Examiner entirely ignores steps (a), (d), (e), (f), and (h) in the Step 2B analysis.”
Examiner respectfully disagrees. The formatting of the 101 rejection is such that in Step 2A, Prong 1, Examiner identifies the judicial exceptions (mental processes/mathematical concepts/methods of organizing human activity) in the claim. In Step 2A, Prong 2, Examiner identifies elements that are not judicial exceptions that may/may not integrate the judicial exception into a practical application. For clarity, Examiner does not rewrite the mental processes in this section. For example, in regards to unsupervised learning, Examiner states that it is simply a high-level recitation of performing unsupervised training to generate a quantum generative model. This should be understood to mean that even in combination with the mental process, it is just a way to “apply it” the mental process using machine learning. Next, for Step 2B, Examiner carries over the identification of additional elements in the claim from Step 2A Prong 2 and re-evaluates any additional element or combination of elements that were considered to be insignificant extra-solution activity and whether those are beyond well-understood, routine, conventional activity. Therefore, Examiner did not omit any steps. All elements were considered individually and as a whole.
Applicant argues (pg. 10): “Examiner’s truncation of claim 1 has effectively reduced the claimed method to a mere abstraction: generating a quantum model, using the quantum model, and repeating the process”, failing to “address the specific technical sequence of the hybrid classical-quantum pipeline.” Applicant goes on to state that by allegedly missing steps c, d, e, and f, the Examiner “fails to account for the specific technical process of filtering samples based on properties and evaluating them through a cost function”
Examiner respectfully disagrees. Examiner did not miss any parts (see the response to the first bullet point above).
Applicant argues (pg. 10) that Examiner’s characterization of “quantum generative model”, “unsupervised learning”, “cost function” are black boxes is legally misplaced in Step 2B and should be addressed under 35 U.S.C. 112.
Examiner respectfully disagrees. This is not a 35 U.S.C. 112 rejection because the elements are not indefinite. By characterizing “quantum generative model” and “unsupervised learning” as black boxes, Examiner means that these additional elements do not, in combination with the judicial exception, integrate it into a practical application than beyond just using a machine learning model or using machine learning training to apply a mental process.
Applicant argues (pg. 10-11) that an element cannot logically simultaneously be a black box and well-understood, routine, and conventional.
Examiner respectfully disagrees. The elements that were claimed to be a black box are not the same elements that were claimed to be well-understood, routine, and conventional. They are different additional elements (for instance, using unsupervised learning is a black box but the step of repeating the loop is WURC, two different elements) and neither integrate into practical application or is significantly more or amounting to inventive concept.
Applicant argues (pg. 11) that “If something is an unknown black box, it cannot, by definition, comprise part of a process that is well- understood enough to be performed entirely within the human mind”, creating a “legal paradox”.
Examiner respectfully disagrees. The elements that were claimed to be a black box are not the same elements that were claimed to be well-understood enough to be mental process. They are different additional elements (for instance, using unsupervised learning is a black box but generating a dataset is a mental process, two different elements) and neither integrate into practical application or is significantly more or amounting to inventive concept.
Applicant argues (pg. 11-12) that in regards to Step 2A, Prong 1, Examiner omits the operative portion of step (c) of using a quantum generative model. Applicant states: “By failing to explain how a human mind can simulate a quantum generative model to produce the recited bit string samples, the Examiner has failed to establish that the claim is directed to a recognized judicial exception.”
Examiner respectfully disagrees. See response to first bullet to see how Examiner formatted the office action. In this particular example, Examiner states that the quantum generative model just amounts to the words “apply it” to merely implement generating bitstring samples. A human can mentally generate bitstring samples and using a generative model is merely a way to apply that, not going beyond the judicial exception. Examiner suggests explaining why using a quantum generative model to produce the bit string samples is necessary for the invention (what about the bitstrings from the quantum model is different from human produced bitstrings and why is that important?)
Applicant argues (pg. 12) that Examiner entirely omits step (b) from Step 2A, Prong 1 and that the Examiner did not consider the claim as a whole. Applicant argues that unsupervised learning cannot be a mental process.
Examiner respectfully disagrees. Please see above for how office action is formatted. In Step 2A, Prong 1, just the elements of judicial exceptions are listed (mental processes). Examiner never said that unsupervised learning is a mental process – Examiner asserts that it is mere instruction to apply the judicial exception (see Prong 2).
Applicant argues (pg. 12-13) that under Step 2A, Prong 2, the ordered combination of filtering samples addresses the technical problem of optimizing generative models that operate on quantum bit string distributions – a technical challenge rooted in quantum data structures that doesn’t exist in the mental realm. Applicant argues that in this step, Examiner does not even consider steps (a), (d), (e), and (f).
Examiner respectfully disagrees. None of the elements in claim 1 reflect optimizing generative models that operate on quantum bit string distributions. Optimizing generative models means that there is something being done in the learning that is making the model better. The closest elements the Examiner sees is filtering bit string samples and applying a cost function to the samples after the training is already done (therefore it seems to be more of a regular function rather than a cost function strictly during training). Therefore, the improvement is more in the cleaning of the data input to the generative model in subsequent iterations rather than some optimization in the generative model itself. In regards to the assertion that Examiner doesn’t consider steps (a), (d), (e), and (f) please see response to first bullet point regarding formatting of 101 rejections.
Claim Rejections – 35 U.S.C. 103
Applicant’s arguments have been fully considered but they are not persuasive.
Applicant argues (pg. 13-5) that Examiner’s definition of filtering is legally and technically deficient, stating that Examiner’s interpretation of any process that restricts is so broad that is would encompass virtually every physical and logical constraint, rendering the term meaningless. Instead, Applicant advocates for filtering as a device or program that separates data, signals, or material according to specified criteria. Applicant argues that Han’s sampling method brings bit string samples into existence, not restricting/separating/reducing. Applicant also argues that logically, you cannot filter bitstrings that doesn’t exist yet (cannot teach generate then filter)
Examiner only partially agrees. Examiner agrees that the definition of filtering is more clearly defined as separating, which was actually what Examiner meant by restricting (i.e. restricting access to a location is separating who can enter vs who cannot). This is a broader definition than the filtering definition of reducing by the Applicant. However, Examiner respectfully disagrees on the remaining points. In regards to Hans’ sampling, Han assembles samples, bit by bit, in a probabilistic manner (supported in Han, Page 3, col. 2). Since this process is probabilistic, the sample created is one bitstring from a possible set of bitstrings that instead could have been created, based on the probability of each bit. Therefore, in effect, Han’s process is indeed a reductive filtering process, as using probability to select each bit in the bitstring chooses a bit (1 or 0) in each bit of the sample string (and the resulting bitstring is just one of the many bitstrings in the probability space of all possible bitstrings, and thus is reductive from the space of many possibilities). Furthermore, as the claim language simply recites filtering the new bit string samples, the process in Han does indeed separate the bit string samples, by the probability function. Since the probability function is built using the properties of the bitstrings, the filtering is based on the properties of the bitstring. To put more clearly, the generate then filter is actually taught by Han because from the perspective of all possible long bitstrings that could be generated (the fact that it’s one-by-one does not matter because the total probability can still be calculated in a chain product), the generation and filtering is done by both (selecting one bitstring from space of many different possible bitstrings.)
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 1-37 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1-17 are method claims. Claims 18-37 are machine/system/product claims. Therefore, claims 1-37 are directed to either a process, machine, manufacture or composition of matter.
With respect to claim 1:
Step 2A – Prong 1:
…
(a) generating a first dataset, the first dataset comprising a plurality of bit string samples from a prior probability distribution, (mental process – a person can manually generate a dataset from probability distribution with the assistance of a pen/paper.)
…
… to generate a plurality of new bit string samples; (mental process – a person can manually generate a plurality of new bit string samples with the assistance of a pen/paper.)
(d) filtering the new bit string samples according to properties of the plurality of new bit string samples to produce a plurality of filtered bit string samples; (mental process – a person can manually filter bit strings with the assistance of a pen/paper.)
(e) applying a cost function to the plurality of filtered bit string samples to produce a plurality of cost function values of the plurality of filtered bit string samples; (mental process – a person can manually apply a cost function to filtered bit strings with the assistance of a pen/paper.)
(f) evaluating the plurality of filtered bit string samples based on the plurality of cost function values of the plurality of filtered bit string samples; (mental process – a person can manually evaluate bit strings based on cost function values with the assistance of a pen/paper.)
(g) selecting a subset of the plurality of filtered bit string samples based on the evaluation; (mental process – a person can manually select a subset of filtered bit string samples with the assistance of a pen/paper.)
(h) merging the first dataset with the subset of the plurality of filtered bit string samples to generate a second dataset; (mental process – a person can manually merge the first dataset with filtered bit string samples to create a second dataset with the assistance of a pen/paper.)
…
Step 2A – Prong 2: This judicial exception is not integrated into a practical application.
A method performed by a computer system for solving combinatorial optimization problems, the computer system comprising a classical computer, the classical computer including a processor, a non-transitory computer-readable medium, and computer instructions stored in the non-transitory computer-readable medium, the computer instructions being executable by the processor to perform the method, the method comprising: (mere instructions to apply the exception using a generic computer component – classical computer, processor, non-transitory computer-readable medium, computer instructions apply exception)
…
(b) performing unsupervised training using the first dataset to generate a quantum generative model; (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 uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: High level recitation of performing unsupervised training to generate a quantum generative model.);
(c) using the quantum generative model … (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 uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: High level recitation of training a quantum generative model to generate bit string samples.);
…
…
…
…
…
and (g) iteratively repeating (c) through (h), wherein in each iteration the output of (h) provides the input to (c) until reaching a limiting number of iterations. (Adding insignificant extra-solution activity to the judicial exception - see MPEP 2106.05(g)).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
A method performed by a computer system for solving combinatorial optimization problems, the computer system comprising a classical computer, the classical computer including a processor, a non-transitory computer-readable medium, and computer instructions stored in the non-transitory computer-readable medium, the computer instructions being executable by the processor to perform the method, the method comprising: (mere instructions to apply the exception using a generic computer component – classical computer, processor, non-transitory computer-readable medium, computer instructions apply exception)
…
(b) performing unsupervised training using the first dataset to generate a quantum generative model; (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 uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: High level recitation of performing unsupervised training to generate a quantum generative model.);
(c) using the quantum generative model … (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 uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: High level recitation of training a quantum generative model to generate bit string samples.);
…
…
…
…
…
and (g) iteratively repeating (c) through (h), wherein in each iteration the output of (h) provides the input to (c) until reaching a limiting number of iterations. (MPEP 2106.05(d)(II) indicate that merely “Performing repetitive calculations” is a well‐understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim – the calculations of the quantum generative model are merely repeated). Thereby, a conclusion that the claimed distribute step is well-understood, routine, conventional activity is supported under Berkheimer.)
With respect to claim 2:
Step 2A – Prong 1:
The method of claim 1, wherein the properties of the plurality of new bit string samples include cardinality constraints. (mental process – a person can recognize that the bit string samples include cardinality constraints).
With respect to claim 3:
Step 2A – Prong 1:
The method of claim 1, wherein the properties of the plurality of new bit string samples include frequency of appearance. (mental process – a person can recognize that the bit string samples include a property of frequency of appearance).
With respect to claim 4:
Step 2A – Prong 1:
The method of claim 1, wherein the prior probability distribution comprises initial observations and cost function values. (mental process – a person can recognize that the prior probability distribution comprises initial observations and cost function values).
With respect to claim 5:
Step 2A – Prong 1:
The method of claim 4, further comprising drawing the initial observations from randomly selected data elements in the first dataset. (mental process – a person can manually draw initial observations from randomly selected data elements in the first dataset with the assistance of a pen/paper.)
With respect to claim 6:
Step 2A – Prong 1:
The method of claim 1, (mental process from claim 1)
Step 2A – Prong 2: This judicial exception is not integrated into a practical application.
wherein (b) comprises using matrix product states (MPS) to generate the quantum generative model. (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 uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: High level recitation using MPS to generate a quantum generative model.);
With respect to claim 7:
Step 2A – Prong 1:
The method of claim 1, wherein the quantum generative model is implemented as a tensor network (TN). (mental process – a person can recognize that the quantum generative model is implemented as a tensor network).
With respect to claim 8:
Step 2A – Prong 1:
The method of claim 1, wherein the quantum generative model comprises a generative adversarial network (GAN). (mental process – a person can recognize that the quantum generative model comprises a generative adversarial network).
With respect to claim 9:
Step 2A – Prong 1:
The method of claim 1, wherein the evaluating comprises evaluating the plurality of filtered bit string samples based on minimizing cost function values. (mental process – a person can recognize that the evaluating comprises evaluating the plurality of filtered bit string samples based on minimizing cost function values).
With respect to claim 10:
Step 2A – Prong 1:
The method of claim 1, wherein the method is practiced in a stand-alone mode. (mental process – a person can recognize that the method is practiced in a stand-alone mode).
With respect to claim 11:
Step 2A – Prong 1:
The method of claim 10, wherein the required number of cost function evaluations is smaller than that of classical optimizers. (mental process – a person can recognize that the required number of cost function evaluations is smaller than that of classical optimizers).
With respect to claim 12:
Step 2A – Prong 1:
The method of claim 1, … and wherein the method boosts performance of the first optimizer. (mental process – a person can recognize that the method boosts performance of the first optimizer).
Step 2A – Prong 2: This judicial exception is not integrated into a practical application.
… wherein (a) comprises receiving the first dataset from an output of a first optimizer, … (Adding insignificant extra-solution activity to the judicial exception - see MPEP 2106.05(g)).
Step 2B: The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
… wherein (a) comprises receiving the first dataset from an output of a first optimizer, … (MPEP 2106.05(d)(II) indicate that merely “Receiving or transmitting data over a network, e.g., using the Internet to gather data” is a well‐understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim – the first dataset from the output of a first optimizer is merely received). Thereby, a conclusion that the claimed distribute step is well-understood, routine, conventional activity is supported under Berkheimer.)
With respect to claim 13:
Step 2A – Prong 1:
The method of claim 13, wherein the method achieves lower minima of the cost function than the first optimizer. (mental process – a person can recognize that the method achieves lower minima of the cost function than the first optimizer).
With respect to claim 14:
Step 2A – Prong 1:
The method of claim 13, wherein the first optimizer comprises a classical optimizer. (mental process – a person can recognize that the first optimizer comprises a classical optimizer).
With respect to claim 15:
Step 2A – Prong 1:
The method of claim 1, wherein the computer system further comprises a quantum computer, the quantum computer comprising a plurality of qubits. (mental process – a person can recognize that the computer system further comprises a quantum computer, the quantum computer comprising a plurality of qubits).
With respect to claim 16:
Step 2A – Prong 1:
The method of claim 15, wherein performing unsupervised training using the first dataset to generate the quantum generative model comprises performing the unsupervised training on the quantum computer. (mental process – a person can recognize that performing unsupervised training using the first dataset to generate the quantum generative model comprises performing the unsupervised training on the quantum computer).
With respect to claim 17:
Step 2A – Prong 1:
The method of claim 15, wherein the quantum generative model comprises a quantum- assisted generative adversarial network (qa-GAN). (mental process – a person can recognize that the quantum generative model comprises a quantum- assisted generative adversarial network).
Claims 18, 19, 20, 21 are rejected on the same grounds under 35 U.S.C. 101 as claims 1, 4, 2, 3, as they are substantially similar, respectively. Mutatis mutandis.
With respect to claim 22:
Step 2A – Prong 1:
The system of claim 18, wherein the method further comprises, before (b), performing initial cost function evaluations on a randomly selected data element in the first dataset. (mental process – a person can recognize that before (b), the method further comprises performing initial cost function evaluations on a randomly selected data element in the first dataset).
Claims 23-31 are rejected on the same grounds under 35 U.S.C. 101 as claims 6-14, as they are substantially similar, respectively. Mutatis mutandis.
With respect to claim 32:
Step 2A – Prong 1:
The system 18, further comprising a Quantum Circuit Associative Adversarial Network (QC-AAN). (mental process – a person can recognize that the system comprises a Quantum Circuit Associative Adversarial Network).
With respect to claim 33:
Step 2A – Prong 1:
The system of claim 18, further comprising a Quantum Circuit Born Machine (QCBM). (mental process – a person can recognize that the system comprises a Quantum Circuit Born Machine).
With respect to claim 34:
Step 2A – Prong 1:
The system of claim 18, wherein the quantum generative model is implemented using gate- based quantum circuits. (mental process – a person can recognize that the quantum generative model is implemented using gate- based quantum circuits).
Claims 35-37 are rejected on the same grounds under 35 U.S.C. 101 as claims 15-17, as they are substantially similar, respectively. Mutatis mutandis.
Claim Rejections – 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-7, 9-10, 18-24, 26-27, 33 are rejected under 35 U.S.C. 103 as being unpatentable over Han et al. (“Unsupervised Generative Modeling Using Matrix Product States”) hereinafter known as Han in view of Yang et al. (“Loop Optimization for Tensor Network Renormalization”) hereinafter known as Yang.
Regarding independent claim 1, Han teaches:
… a non-transitory computer-readable medium, … (Han [Page 7, Column 2, Paragraph 1]: “Fortunately, Fig. 2(b) also shows that the decay of memory capability with system size can be compensated by increasing D_{max}” Han teaches a computer-readable storage medium in the form of memory. Han teaches that the memory usage can be compensated for by increasing a parameter in the model.)
(a) generating a first dataset, the first dataset comprising a plurality of bit string samples from a prior probability distribution, (Han [Page 2, Column 2, Paragraph 1]: “we consider a data set T consisting of binary strings v ∈ V = {0, 1}^{⊗N}, which are potentially repeated and can be mapped to basis vectors of a Hilbert space of dimension 2^N.” Han teaches that the first dataset of bit string samples is generated from a Hilbert space of dimension 2^N, a distribution of space that the samples can be in probabilistically.)
(b) performing unsupervised training using the first dataset to generate a quantum generative model; (Han [Page 2, Column 1, Paragraph 5]: “The goal of unsupervised generative modeling is to model the joint probability distribution of given data. With the trained model, one can then generate new samples from the learned probability distribution … The probabilistic interpretation of quantum mechanics naturally suggests modeling data distribution with a quantum state” Han teaches modeling data distribution with a quantum state for the purpose for generating new data. To do this, Han teaches using unsupervised training in its quantum generative modeling.)
(c) using the quantum generative model to generate a plurality of new bit string samples; (Han [Page 4, Column 1, Paragraph 1]: “Our model enjoys a direct sampling method, which generates a sample bit by bit from one end of the MPS to the other” Han teaches the MPS, which is a matrix product state model. The MPS generates new bit string samples, bit by bit from one end to the other.)
(d) filtering the new bit string samples according to properties of the plurality of new bit string samples to produce a plurality of filtered bit string samples; (Han [Page 4, Column 1, Paragraph 2]: “Given the value of the Nth bit, one can then move on to sample the (N − 1)th bit.” Han teaches that the bit of the N-1 index is sampled given the value of the bit of the N index. This shows that the new bit string is filtered based on the properties of the new bit string itself.)
…
…
…
(h) merging the first dataset with the subset of the plurality of filtered bit string samples to generate a second dataset; (Han [Page 7, Column 2, Paragraph 3]: “Having chosen |T | = 1000 MNIST images, we train the MPS with different maximal bond dimensions” Han [Page 8, Column 1, Paragraph 2]: “The samples directly generated after training are shown in Fig. 5(a)” Han teaches that the MNIST dataset is used to generate a new set of MNIST images, which are from the filtered bit string samples. When the combination of the original dataset and this new set of MNIST images are considered, then the total is a second dataset of larger size.)
and (g) iteratively repeating (c) through (h), wherein in each iteration the output of (h) provides the input to (c) until reaching a limiting number of iterations. (Han [Page 5, Column 2, Paragraph 6]: “After being trained over four loops of batch gradient descent training, the cost function converges to its minimum value, which is equal to the Shannon entropy of the BS data set.” Han [Page 6, Column 1, Paragraph 4]: “We have generated N_{s} = 10^6 independent samples from the learned MPS. All these samples are training images.” Han teaches that the training occurs in iterative loops by using an example of training that ends in four loops, as the cost function converges to the minimum. Han teaches that the limiting number of iterations is that which converges the cost function to its minimum. Han also teaches that the outputted samples from the MPS are afterward inputted to be training images.)
Han does not explicitly teach:
A method performed by a computer system for solving combinatorial optimization problems, the computer system comprising a classical computer, the classical computer including a processor, … and computer instructions stored in the non-transitory computer-readable medium, the computer instructions being executable by the processor to perform the method, the method comprising:
…
…
…
…
(e) applying a cost function to the plurality of filtered bit string samples to produce a plurality of cost function values of the plurality of filtered bit string samples;
(f) evaluating the plurality of filtered bit string samples based on the plurality of cost function values of the plurality of filtered bit string samples;
(g) selecting a subset of the plurality of filtered bit string samples based on the evaluation;
…
…
However, Yang teaches:
A method performed by a computer system for solving combinatorial optimization problems, the computer system comprising a classical computer, the classical computer including a processor, … and computer instructions stored in the non-transitory computer-readable medium, the computer instructions being executable by the processor to perform the method, the method comprising: (Yang [Page 1, Column 1, Paragraph 1]: “the tensor network (TN) approach has become a powerful theoretical and computational tool for studying condensed matter systems. Many physical quantities, including the partition function of a classical system, the Euclidean path integral of a quantum system, and the expectation value of physical observables, can be expressed in terms of tensor networks. Yang teaches a tensor network approach that is implemented on a computer. This is in a classical computer because the tensor network expresses the classical system. This must be executed by some instructions that are processed on the computer.)
…
…
…
…
(e) applying a cost function to the plurality of filtered bit string samples to produce a plurality of cost function values of the plurality of filtered bit string samples; (Yang [Page 2, Column 2, Paragraph 3]: “Minimizing the cost function is equivalent to minimizing the distance between two MPS. Thus, S tensors can be optimized using the well-developed variational MPS method” Yang teaches that the tensors, which may be bit strings, can be optimized by minimizing the distance between two MPS. As the generating of new data is represented as MPS, this is equivalent to minimizing the cost function to produce a set of bit strings that satisfy this optimization.)
(f) evaluating the plurality of filtered bit string samples based on the plurality of cost function values of the plurality of filtered bit string samples; (Yang [Page 2, Column 2, Paragraph 2]: “In the LN-TNR algorithm, this is achieved by minimizing the following single-site cost functions: The optimal S values are found using SVD and keeping only the largest singular χ values.” Yang teaches that the tensors, which may be bit string samples, are evaluated based on the cost function. The optimal values are found using this cost function and only the largest are singular values are selected.)
(g) selecting a subset of the plurality of filtered bit string samples based on the evaluation; (Yang [Page 2, Column 2, Paragraph 2]: “In the LN-TNR algorithm, this is achieved by minimizing the following single-site cost functions: The optimal S values are found using SVD and keeping only the largest singular χ values.” Yang teaches that the tensors, which may be bit string samples, are evaluated based on the cost function. The optimal values are found using this cost function and only the largest are singular values are selected.)
…
…
Han and Yang are in the same field of endeavor as the present invention, as the
references are directed to generating new data using a quantum generative model and optimization in the context of tensor networks and matrix product states (MPS), respectively. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine generating new data using a quantum generative model as taught in Han with evaluating the new data by minimizing the distances of the MPS of the data as taught in Yang. Yang provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Han to include teachings of Yang because the combination would allow for data generated from MPS to be evaluated by a cost function. This has the potential benefit of only including the data generated from the quantum generative model that satisfies the conditions of the cost function – effectively doing so even though the data was created and was represented using MPS.
Regarding dependent claim 2, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein the properties of the plurality of new bit string samples include cardinality constraints. (Han [Page 4, Column 2, Paragraph 3]: “The bond dimensions of MPS put an upper bound on its ability to capture entanglement entropy.” Han teaches that the bit string samples have an upper bound on them because of the bond dimension of the MPS. As the bond dimension is the dimension of the indices of the bit strings that are generated by the MPS, constraining this bond dimension puts a constraint on the bit string samples.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 3, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein the properties of the plurality of new bit string samples include frequency of appearance. (Han [Page 4, Column 1, Equation 6]: Han teaches that the (k-1)th bit is sampled according to a conditional probability. Since these are bits that can take a value of 0 or 1, this is effectively a frequency of appearance, as the frequency of bits, over a large number of samples, tend to this probability.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 4, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein the prior probability distribution comprises initial observations and cost function values. (Han [Page 2, Column 2, Paragraph 1]: “we consider a data set T consisting of binary strings v ∈ V = {0, 1}^{⊗N}, which are potentially repeated and can be mapped to basis vectors of a Hilbert space of dimension 2^N.” Han teaches that the distribution of possible values is in the Hilbert space of dimension 2^N. The samples of binary strings from this dataset are the initial observations. This data, as it may be input to the cost function to determine the effectiveness of the data generating, are also cost function values.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 5, Han and Yang teach:
The method of claim 4,
Han teaches:
further comprising drawing the initial observations from randomly selected data elements in the first dataset. (Han [Page 2, Column 2, Paragraph 1]: “we consider a data set T consisting of binary strings v ∈ V = {0, 1}^{⊗N}, which are potentially repeated and can be mapped to basis vectors of a Hilbert space of dimension 2^N.” Han teaches that the first dataset of bit string samples is generated from a Hilbert space of dimension 2^N, a distribution of space that the samples can be in probabilistically.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 6, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein (b) comprises using matrix product states (MPS) to generate the quantum generative model. (Han [Page 2, Column 1, Paragraph 3]: “The MPS model also enjoys a direct sampling method [36] much more efficient than that of the Boltzmann machines, which require a Markov chain Monte Carlo (MCMC) process for data generation.” Han teaches that in generating the new data using the quantum generative model, the process includes using a MPS model. Han teaches that the MPS model has direct sampling that is more efficient than other data generation methods, such as Boltzmann machines.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 7, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein the quantum generative model is implemented as a tensor network (TN). (Han [Page 2, Column 1, Paragraph 2]: “In particular, the matrix product state (MPS) is a kind of TN where the tensors are arranged in a one-dimensional geometry” Han teaches that the MPS quantum generative model is a kind of a tensor network (TN). Han specifies that the tensors of the network are arranged in a specific geometry.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 9, Han and Yang teach:
The method of claim 1,
Yang teaches:
wherein the evaluating comprises evaluating the plurality of filtered bit string samples based on minimizing cost function values. (Yang [Page 2, Column 2, Paragraph 3]: “Minimizing the cost function is equivalent to minimizing the distance between two MPS. Thus, S tensors can be optimized using the well-developed variational MPS method” Yang teaches that the tensors, which may be bit strings, can be optimized by minimizing the distance between two MPS. As the generating of new data is represented as MPS, this is equivalent to minimizing the cost function to produce a set of bit strings that satisfy this optimization.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 9, Han and Yang teach:
The method of claim 1,
Yang teaches:
wherein the evaluating comprises evaluating the plurality of filtered bit string samples based on minimizing cost function values. (Yang [Page 2, Column 2, Paragraph 3]: “Minimizing the cost function is equivalent to minimizing the distance between two MPS. Thus, S tensors can be optimized using the well-developed variational MPS method” Yang teaches that the tensors, which may be bit strings, can be optimized by minimizing the distance between two MPS. As the generating of new data is represented as MPS, this is equivalent to minimizing the cost function to produce a set of bit strings that satisfy this optimization.)
The reasons to combine are substantially similar to those of claim 1.
Regarding dependent claim 10, Han and Yang teach:
The method of claim 1,
Han teaches:
wherein the method is practiced in a stand-alone mode. (Han [Page 8, Column 2, Paragraph 3]: “In a glimpse of its generalization ability, we also tried reconstructing MNIST images other than the training images, as shown in Figs. 6(c) and 6(d). These results indicate that the MPS has learned crucial features of the data set, rather than merely memorizing the training instances” Han teaches that the model is used, self-sufficiently, on the MNIST images dataset. The fact that the model, after training, can be used on unrelated data shows that the model is able to be practiced in a stand-alone mode, without the need for the data to be managed to fit a specific use case.)
The reasons to combine are substantially similar to those of claim 1.
Claims 18, 19, 20, 21 are rejected on the same grounds under 35 U.S.C. 103 as claims 1, 4, 2, 3, as they are substantially similar, respectively. Mutatis mutandis.
Regarding dependent claim 22, Han and Yang teach:
The system of claim 18,
Han teaches:
wherein the method further comprises, before (b), performing initial cost function evaluations on a randomly selected data element in the first dataset. (Han [Page 3, Column 1, Paragraph 5]: “Minimizing the NLL reduces the dissimilarity between the model probability distribution P(v) and the empirical distribution defined by the training set” Han teaches that the NLL (negative log likelihood) cost function is evaluated on the training set that is distributed by a probability distribution. This shows that the elements in the training dataset is chosen probabilistically, or randomly.)
The reasons to combine are substantially similar to those of claim 1.
Claims 23, 24, 26, 27 are rejected on the same grounds under 35 U.S.C. 103 as claims 6, 7, 9, 10, as they are substantially similar, respectively. Mutatis mutandis.
Regarding dependent claim 33, Han and Yang teach:
The system of claim 18,
Han teaches:
further comprising a Quantum Circuit Born Machine (QCBM). (Han [Page 1, Column 2, Paragraph 3]: “Precisely speaking, it is the wave functions that are modeled in quantum physics, and probability distributions are given by their squared norm according to Born’s statistical interpretation. … Hence, we may refer to probability models that exploit quantum state representations as ‘Born machines’” Han teaches that the system further comprises a born machine as a probability model to represent quantum states. This is also known as a Quantum Circuit Born Machine.)
The reasons to combine are substantially similar to those of claim 1.
Claims 8, 11-17, 25, 28-32, 34-37 are rejected under 35 U.S.C. 103 as being unpatentable over Han in view of Yang in view of Dallaire-Demers et al. (“Quantum generative adversarial networks”) hereinafter known as Dallaire.
Regarding dependent claim 8, Han and Yang teach:
The method of claim 1,
Han and Yang do not explicitly teach:
wherein the quantum generative model comprises a generative adversarial network (GAN).
However, Dallaire teaches:
wherein the quantum generative model comprises a generative adversarial network (GAN). (Dallaire [Page 2, Column 2, Paragraph 2]: “The general aim of training a GAN is to find a generator G which mimics the real data source R. In the quantum case, we define G to be a variational quantum circuit whose gates are parametrized by a vector” Dallaire teaches that the generation for data can be done using a GAN. Dallaire teaches that the GAN can be used in the quantum case as well.)
Dallaire is in the same field as the present invention, since it is directed to using a generative adversarial network (GAN) to generate a new dataset, which includes a discriminator that distinguishes fake data from real data. It would have been obvious, before the effective filing date of the claimed invention, to a person of ordinary skill in the art, to combine a quantum generative model to generate new data as taught in Han as modified by Yang with specifically using a generative adversarial network as taught in Dallaire. Dallaire provides this additional functionality. As such, it would have been obvious to one of ordinary skill in the art to modify the teachings of Han as modified by Yang to include teachings of Dallaire because the combination would allow for the generation of new data to be assessed based on how real or fake it is using the discriminator, part of the GAN. This has the potential benefit of generating new data that is as close to the real data as possible, as measured by the GAN. This allows for the size of the original dataset to be increased with samples that are largely indistinguishable from real data, which may help in training other models that require this informative data.
Regarding dependent claim 11, Han and Yang teach:
The method of claim 10,
Dallaire teaches:
wherein the required number of cost function evaluations is smaller than that of classical optimizers. (Dallaire [Page 3, Column 1, Paragraph 4]: “For classical GANs, the optimization task is traditionally defined with log-likelihood functions but it is more convenient to define a cost function linear in the output probabilities of D in the quantum case since we want to optimize a function which is linear in some expectation value.” Dallaire teaches that the classical GAN is defined with multiple log-likelihood functions but the quantum case requires just one cost function linear to the output probabilities.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 12, Han and Yang teach:
The method of claim 1,
Dallaire teaches:
wherein (a) comprises receiving the first dataset from an output of a first optimizer, and wherein the method boosts performance of the first optimizer. (Dallaire [Page 2, Column 1, Paragraph 2]: “G transforms this noise source into data samples … creating the generator distribution pG(x). … the task of training G corresponds to the task of maximizing the probability that D misclassifies a generated sample as an element of the real data.” Dallaire teaches that a dataset may be output from a generator that maximizes the probability that one cannot tell the difference between the generated dataset and the real data (in other words, maximizing the misclassification).
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 13, Han, Yang, and Dallaire teach:
The method of claim 12,
Dallaire teaches:
wherein the method achieves lower minima of the cost function than the first optimizer. (Dallaire [Page 8, Column 2, Paragraph 1]: “We have reformulated the optimization problem of GANs in the quantum formalism, yielding QuGANs. We have shown how the cost function can be optimized by directly evaluating the gradients with a quantum processor.” Dallaire teaches that the first optimizer was reformulated to the quantum formalization. Dallaire teaches that the cost function can be optimized by the quantum method. This means that the method achieves at least as low or a minima, or lower, of the cost function as it is optimized.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 14, Han, Yang, and Dallaire teach:
The method of claim 13,
Dallaire teaches:
wherein the first optimizer comprises a classical optimizer. (Dallaire [Page 2, Column 1, Paragraph 2]: “G transforms this noise source into data samples … creating the generator distribution pG(x). … the task of training G corresponds to the task of maximizing the probability that D misclassifies a generated sample as an element of the real data.” Dallaire teaches that a dataset may be output from a generator that maximizes the probability that one cannot tell the difference between the generated dataset and the real data (in other words, maximizing the misclassification). This optimization is a classical optimizer, as this instance is without the quantum case.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 15, Han and Yang teach:
The method of claim 1,
Dallaire teaches:
wherein the computer system further comprises a quantum computer, the quantum computer comprising a plurality of qubits. (Dallaire [Page 1, Column 2, Paragraph 2]: “Quantum computers have the potential to solve problems believed to be beyond the reach of classical computers, such as factoring large integers.” Dallaire [Page 7, Column 1, Paragraph 1]: “Since the generators of those gates are all simple Pauli operators, it is easy to implement the conditional hj ’s with CNOTs, CPHASEs and CZZs where the ZZs are between nearest-neighbor qubits.” Dallaire teaches a quantum computer to solve problems that classical computers have difficulty solving. The quantum computer is made up of gates that comprise of qubits.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 16, Han, Yang, and Dallaire teach:
The method of claim 15,
Han teaches:
wherein performing unsupervised training using the first dataset to generate the quantum generative model comprises performing the unsupervised training on the quantum computer. (Han [Page 2, Column 1, Paragraph 5]: “The goal of unsupervised generative modeling is to model the joint probability distribution of given data. With the trained model, one can then generate new samples from the learned probability distribution … The probabilistic interpretation of quantum mechanics naturally suggests modeling data distribution with a quantum state” Han teaches modeling data distribution with a quantum state for the purpose for generating new data. To do this, Han teaches using unsupervised training in its quantum generative modeling, which is done on the quantum computer.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 17, Han, Yang, and Dallaire teach:
The method of claim 15,
Dallaire teaches:
wherein the quantum generative model comprises a quantum- assisted generative adversarial network (qa-GAN). (Dallaire [Page 2, Column 1, Paragraph 3]: “We will formalize QuGANs as a quantum generalization of conditional GANs… A well-trained QuGAN could produce new molecular states which also have the same properties but were not in the original dataset. In another context, a QuGAN could be used to compress time evolution gate sequences for different time steps to use in larger quantum simulations.” Dallaire teaches that the model comprises of GANs that are different to be a quantum generalization. This is called a QuGAN.)
The reasons to combine are substantially similar to those of claim 8.
Claims 25, 28, 29, 30, 31 are rejected on the same grounds under 35 U.S.C. 103 as claims 8, 11, 12, 13, 14 as they are substantially similar, respectively. Mutatis mutandis.
Regarding dependent claim 32, Han and Yang teach:
The system 18,
Dallaire teaches:
further comprising a Quantum Circuit Associative Adversarial Network (QC-AAN). (Dallaire [Page 7, Column 1, Paragraph 1]: “Since the generators of those gates are all simple Pauli operators, it is easy to implement the conditional hj ’s with CNOTs, CPHASEs and CZZs where the ZZs are between nearest-neighbor qubits.” Dallaire teaches a quantum generative model to generate new data. The quantum generative model is made up of gates that comprise of qubits. As the QC-AAN is a network that generates new data using the quantum model comprised of qubits, Dallaire teaches that the system is comprised of a QC-AAN.)
The reasons to combine are substantially similar to those of claim 8.
Regarding dependent claim 34, Han and Yang teach:
The system of claim 18,
Dallaire teaches:
wherein the quantum generative model is implemented using gate- based quantum circuits. (Dallaire [Page 7, Column 1, Paragraph 1]: “Since the generators of those gates are all simple Pauli operators, it is easy to implement the conditional hj ’s with CNOTs, CPHASEs and CZZs where the ZZs are between nearest-neighbor qubits.” Dallaire teaches a quantum generative model to generate new data. The quantum generative model is made up of gates that comprise of qubits.)
The reasons to combine are substantially similar to those of claim 8.
Claims 35-37 are rejected on the same grounds under 35 U.S.C. 103 as claims 15-17 as they are substantially similar, respectively. Mutatis mutandis.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to KYU HYUNG HAN whose telephone number is (703) 756-5529. The examiner can normally be reached on MF 9-5.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached on (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.
/Kyu Hyung Han/
Examiner
Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights