DETAILED ACTION
Application No. 18/388,115 filed on 11/08/2023 has been examined. In this Office Action, claims 1-18 are pending.
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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 11/08/2023 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
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 17-18 are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter.
Claim 17 is rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter. In view of applicant specification it is not clear if system include definitive hardware or physical components. Applicant is suggested to insert – “memory and processor” in the claims to obviate this rejection.
Claim 18 is also rejected under 35 U.S. C 101 because they fail to resolve the deficiencies of claim 17.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 10-11 and 13-18 are rejected under 35 U.S.C. §103 as being unpatentable over Chen et al., “Hybrid quantum-classical classifier based on tensor network and variational quantum circuit,” arXiv:2011.14651v1, Nov. 30, 2020, in view of Arthur et al., “A Hybrid Quantum-Classical Neural Network Architecture for Binary Classification,” arXiv:2201.01820v2, Jan. 11, 2022, and further in view of Orus Lacort et al (US 2023/0367822 A1).
As per claim 1, the method includes training a hybrid quantum-classical computation system for approximating a labeling function for an input feature vector, the system comprising (Chen teaches a hybrid quantum-classical classifier based on a tensor network and a variational quantum circuit for supervised learning and binary classification, Contracting the feature-mapped input and the MPS yields a four-dimensional feature vector, which corresponds to a compressed representation to be used as an input for the VQC to perform classification. Chen, p. 1, Abstract; p. 3, Section 2.3; p. 4, Fig. 3.): a variational quantum circuit comprising a plurality of quantum gates acting on qubits of a qubit register, the plurality of quantum gates comprising variational quantum gates, wherein parametrized actions of the variational quantum gates on the qubits of the qubit register is parametrized according to associated variational parameters, and at least one encoding gate for modifying a state of the qubits of the qubit register according to the input feature vector (Chen teaches a VQC having an encoding block U(x) and a variational block Φ(θ), and teaches parameterized gates R(α, β, γ) in the VQC. Chen, p. 3, Section 2.2; p. 4, Fig. 4, further see encoding gates Ry(arctan(xi)) and Rz(arctan(xi²)) for state preparation. Chen, p. 4, Fig. 4.);
a machine learning model, implemented on a classical processing system, configured to process the input feature vector according to a parametrized transfer function (Chen teaches that the tensor network/MPS is used as a classical feature extractor that processes input data before the VQC. Chen, p. 2, Section 2.1; p. 3, Section 2.3), wherein the parametrized transfer function is parametrized by machine- learning parameters (Chen teaches that the MPS contains trainable parameters. Chen, p. 4, Fig. 3.);
a labeling module, implemented on a classical processing system, configured to receive a first output generated by the variational quantum circuit and a second output generated by the machine learning model and to generate an output label based on a combination of the first output and the second output, wherein the combination is based on a plurality of trainable combination parameters (Chen teaches a hybrid TN-VQC classifier in which the classical tensor network output is provided to the VQC for classification, and the first two qubits are measured for classification labels. Chen, p. 3, Section 2.3; p. 4, Fig. 3 and Fig. 4.);
The method includes an iterative training process comprising the steps of: providing an input feature vector of the sample dataset to the variational quantum circuit and to the machine learning model (Chen teaches that the hybrid TN-VQC architecture is trained end-to-end, and that parameters within the MPS are updated together with parameters within the VQC at each iteration. Chen, p. 4, paragraph following Fig. 4.);
Chen does not explicitly teach the machine- learning parameters and the trainable combination parameters based on a value of a cost function.
However, Arthur teaches the machine- learning parameters and the trainable combination parameters based on a value of a cost function (e.g., teaches that a VQC includes a feature map F, an ansatz A(θ), and an observable measurement producing an expectation value. Arthur, p. 2, Section II, Equations 1–3, minimizing a negative log-likelihood cost function and computing derivatives with respect to VQC parameters. Arthur, pp. 2–3, Section II, Equations 6–8, further teaches parameter-shift-rule gradient calculation and backpropagation for hidden-layer and output-layer parameters. Arthur, p. 3, Section II, Equations 8–12).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the teachings of Arthur with the teachings of Chen in order to improve training of a hybrid quantum-classical classifier by using known variational quantum circuit training techniques, including cost-function minimization, gradient calculation, parameter-shift derivatives, and backpropagation, as taught by Arthur.
Chen and Arthur do not explicitly teach providing the first output and the second output to the labeling module, and determining a parameter update of the variational parameters based on a value of a cost function for the output label for the input feature vector.
However, Orus teaches providing the first output and the second output to the labeling module, and determining a parameter update of the variational parameters based on a value of a cost function for the output label for the input feature vector (e.g., an initial cost function calculation step wherein the classical digital computer calculates the cost of the cost function based on the initial position vector of each data point of the data set, wherein the classical digital computer executes a classical optimizer for the cost function, obtaining a set of optimized variational parameters and an optimized variational parameter, and the initial label vector of each of the labels on the Bloch sphere, sending a final quantum state to a classical digital computer and calculating a cost function based on the final quantum state and labels, ¶[0018]–¶[0029], ¶[0073]–¶[0074]).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the teachings of Orus with the combined teachings of Chen and Arthur in order to improve hybrid classical-quantum optimization by using a classical optimizer to calculate a cost function based on a final quantum state and labels, thereby providing a known way to evaluate and optimize quantum outputs in a hybrid quantum-classical system, as taught by Orus.
As per claim 10, determining the parameter update comprises determining a vector of derivatives for the variational parameters as part of a parameter update gradient.
(Chen teaches VQC variational parameters updated during end-to-end training. Chen, p. 4, Fig. 4 and paragraph following Fig. 4.).
Chen does not explicitly teach determining a vector of derivatives for the variational parameters.
However, Arthur teaches determining a vector of derivatives for the variational parameters (e.g., determining derivatives for VQC variational parameters. Arthur, p. 3, Section II, Equation 8, and computing derivatives of the VQC expectation value with respect to each ansatz parameter. Arthur, p. 3, Section II, Equation 8).
As per claim 11, determining the vector of derivatives comprises applying the parameter-shift rule to a subset of or all of the variational gates at each iteration.
(Chen teaches a VQC having variational parameters. Chen, p. 4, Fig. 4).
Chen does not explicitly teach applying the parameter-shift rule.
However, Arthur teaches applying the parameter-shift rule (e.g., teaches applying the parameter-shift rule to compute derivatives for variational parameters. Arthur, p. 3, Section II, Equation 8. and further teaches using the parameter-shift rule to compute the gradient of each circuit during hybrid neural-network training. Arthur, p. 3, Section II, paragraph following Equation 12).
As per claim 13, wherein the at least one encoding gate comprises single-qubit rotations proportional to a value of the input feature vector (Chen teaches encoding gates Ry(arctan(xi)) and Rz(arctan(xi²)) for state preparation. Chen, p. 4, Fig. 4, and Orus also teaches locating each data point on the Bloch sphere by rotating in a Y direction and a Z direction, where each angle is proportional to the value of a corresponding attribute. Orus, ¶[0062], ¶[0116]).
As per claim 14, wherein the at least one encoding gate is applied k times as part of the variational quantum circuit. (Chen teaches a VQC including encoding and variational blocks. Chen, p. 4, Fig. 4.), and Orus teaches applying the encoding gate k times (Orus teaches that the quantum circuit comprises as many layers on each qubit as the number of groups of three variational parameters. Orus, ¶[0068], ¶[0109], ¶[0125]).
As per claim 15, k is an integer value greater than 2 (Orus teaches that the quantum circuit may comprise more than one group of three variational parameters and sequential rotations for each group. Orus, ¶[0120].), and examples of quantum gates 11, 12, and 13, each corresponding to a group of three optimized variational parameters. Orus, ¶[0125]).
As per claim 16, wherein the variational quantum circuit is parametrized by at least 2k variational parameters (Chen teaches parameterized variational gates in the VQC. Chen, p. 4, Fig. 4, and that each group/layer includes three variational parameters, that the number of layers corresponds to the number of groups of three variational parameters, ¶[0068], ¶[0070], ¶[0125]).
Regarding claim 17, claim 17 is rejected for substantially the same reason as claim 1 above.
As per claim 18, wherein the combination parameters determine a ratio of a relative contribution of the first output and the second output to the output label, wherein the ratio is greater than 0.01 (Chen Fig,4, teaches wherein The CNOT gates are used to entangle quantum states from each qubit and R(α,β,γ) represents the general single qubit unitary gate with three learnable parameters αi, βi and γi. The first two qubits are measured for classification labels and Figure 6: Result: MPS classifier. Results of the binary classification task using an MPS as a classifier. Bond dimension of the MPS is (a) χ = 1 (b) χ = 2, the results of the MPS classifier for χ = 1 and 2. For χ = 1, the accuracy of both the training and testing datasets remain around 68 − 70%. When we increase the bond dimension to χ = 2, we observe the accuracy reaching close to 99%).
Claims 2-8 are rejected under 35 U.S.C. §103 as being unpatentable over Chen et al. in view of Arthur et al., in view of Orus Lacort et al (US 2023/0367822 A1), further in view of Kingma et al., “Adam: A Method for Stochastic Optimization,” arXiv 2014 paper.1412.6980.pdf, dated December 22, 2014.
As per claim 2, wherein a learning rate for updating the variational parameters and the machine-learning parameters is different (Chen teaches separate trainable parameter groups for the classical MPS and the VQC because the MPS parameters are updated together with the VQC parameters during end-to-end training. Chen, p. 4, paragraph following Fig. 4.).
Chen does not explicitly teach different learning rates for the variational parameters and the machine-learning parameters.
However, Adam/Kingma teaches different learning rates for the variational parameters and the machine-learning parameters (e.g., which computes individual adaptive learning rates for different parameters. Kingma, p. 1, Introduction and an element-wise adaptive parameter update using α · m̂t/(√v̂t + ε). Kingma, p. 2, Algorithm 1.).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to apply the teachings of Adam/Kingma with the teachings of Chen, Arthur and Orus in order to improve updating of multiple trainable parameter groups in a hybrid quantum-classical model by using adaptive learning rates for different parameters, as taught by Adam/Kingma.
As per claim 3, the method determines an update vector for the variational parameters, the machine-learning parameters, and the trainable combination parameters (Chen teaches updating MPS parameters and VQC parameters together during end-to-end training. Chen, p. 4, paragraph following Fig. 4.), and wherein determining the parameter update comprises multiplying the update vector with a learning-rate vector comprising different learning-rate factors for the variational parameters and machine-learning parameters and (Adam/Kingma teaches determining a gradient vector gt and updating a parameter vector θ. Kingma, p. 2, Algorithm 1 and teaches that all operations on vectors are element-wise. Kingma, p. 2, Algorithm 1 and further teaches individual adaptive learning rates for different parameters. Kingma, p. 1, Introduction.).
As per claim 4, wherein the update vector is a gradient of the variational parameters, the machine-learning parameters, and the trainable combination parameters with respect to the cost function.
(Chen teaches end-to-end updating of MPS parameters and VQC parameters. Chen, p. 4, paragraph following Fig. 4.), Chen does not explicitly teach the update vector as a gradient with respect to the cost function. However, Arthur teaches the update vector as a gradient with respect to the cost function (Arthur teaches minimizing a cost function and computing gradients with respect to VQC hidden-layer and output-layer parameter vectors. Arthur, pp. 2–3, Section II, Equations 6–12. and Adam/Kingma also teaches determining a gradient vector with respect to a stochastic objective and teaches determining a gradient vector gt = ∇θft(θt−1) with respect to a stochastic objective. Kingma, p. 2, Algorithm 1.).
As per claim 5, determining the parameter update is based on a gradient of the cost function and a learning rate.
(Chen teaches end-to-end training of the MPS-VQC model. Chen, p. 4, paragraph following Fig. 4.). Chen does not explicitly teach that the parameter update is based on both a gradient of the cost function and a learning rate. However, Arthur teaches the gradient portion (Arthur teaches minimizing a cost function and computing derivatives with respect to VQC parameters. Arthur, pp. 2–3, Section II, Equations 6–8.)
Adam/Kingma teaches the learning-rate update portion (Adam/Kingma teaches an update rule based on a gradient and stepsize α. Kingma, p. 2, Algorithm 1.).
As per claim 6, determining the parameter update is based on stochastic gradient descent. Chen and Arthur do not explicitly teach stochastic gradient descent for determining the parameter update. However, Orus teaches stochastic gradient descent for determining the parameter update (Orus teaches that the classical optimizer is a stochastic gradient descent optimizer. Orus, ¶[0079], ¶[0154] and further teaches that the stochastic gradient descent optimizer may be an Adam optimizer or a Momentum optimizer. Orus, ¶[0080]–¶[0081], ¶[0155]–¶[0156].).
As per claim 7, the stochastic gradient descent includes a momentum coefficient based on a previously determined gradient of the cost function.
Chen and Arthur do not explicitly teach a momentum coefficient based on a previously determined gradient. However, Orus teaches use of a Momentum optimizer.
(Orus teaches that the stochastic gradient descent optimizer may be a Momentum optimizer. Orus, ¶[0081], ¶[0156].), Orus does not explicitly provide the detailed momentum coefficient calculation. However, Adam/Kingma teaches provide the detailed momentum coefficient calculation (e.g., teaches updating the first moment estimate as mt ← β1 · mt−1 + (1 − β1) · gt, where mt−1 is based on previous gradient information. Kingma, p. 2, Algorithm 1 and further teaches that β1 controls an exponential moving average over previous gradients. Kingma, p. 2, Section 2.).
As per claim 8, determining the parameter update is based on an update function of a moving average over a gradient of the cost function and a moving average over the squared gradient of the cost function.
Chen and Arthur do not explicitly teach an update function using moving averages of gradients and squared gradients.
However, Orus teaches an update function using moving averages of gradients and squared gradients (e.g., teaches that the stochastic gradient descent optimizer may be an Adam optimizer, ¶[0080], ¶[0155].)
Orus does not explicitly provide moving-average update equations.
However, Adam/Kingma teaches moving-average update equations (e.g., teaches updating mt as a moving average over gradients and updating vt as a moving average over squared gradients. Kingma, p. 2, Algorithm 1, and further teaches using m̂t and v̂t in the update rule θt ← θt−1 − α · m̂t/(√v̂t + ε). Kingma, p. 2, Algorithm 1.).
It is noted that any citation [[s]] to specific, pages, columns, lines, or figures in the prior art references and any interpretation of the references should not be considered to be limiting in any wav. A reference is relevant for all it contains and may be relied upon for all that it would have reasonably suggested to one having ordinary skill in the art. [[See, MPEP 2123]].
Allowable Subject Matter
Claims 9 and 12 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Citation of Pertinent Prior Arts
The prior art made of record and not relied upon in form PTO-892, if any, is considered pertinent to applicant's disclosure.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Mohammad A Sana whose telephone number is (571)270-1753. The examiner can normally be reached Monday-Friday 9-5.
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, Sanjiv Shah can be reached at 5712724098. 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.
/Mohammad A Sana/Primary Examiner, Art Unit 2166
Prosecution Insights