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finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 04/07/2026 has been entered.
Response to Argument
The arguments filed 04/07/2026 have been entered. Claims 1-28 remain pending in the
application.
Applicant’s arguments and amendments, with respect to claim rejections of claims 1-28 under 35 U.S.C 103 filed 04/07/2026 have been considered and some of them are persuasive.
Applicant argues that the cited combination fails to teach or suggest the amended limitations of claim 1. First, Applicant argues that Chudak’s quantum processor does not provide a prior distribution to the GAN generator. According to Applicant, Chudak’s quantum processor provides parameters such as biases h and coupling strengths J, whereas the noise prior distribution r(z) supplied to the generator is separate from the quantum-generated information. Applicant further contends that Liu does not remedy this deficiency because Liu describes a QCBM as a quantum generative model for generating a probability distribution, but does not expressly disclose supplying that distribution as the prior distribution of a separate GAN generator.
Applicant further argues that Donahue does not teach the amended discriminator limitation requiring “an internal latent space defined by activations of a layer of the discriminator.” Applicant contends that the latent variable z discussed by Donahue is an external input supplied to the discriminator, rather than an internal latent space arising from activations of a discriminator layer.
Applicant additionally argues that the proposed combination is based on impermissible hindsight and lacks an adequate motivation to combine the cited references, particularly because the references allegedly do not suggest modifying the GAN so that a quantum-generated probability distribution is supplied as the generator prior or configuring the discriminator in the manner now claimed.
Applicant’s argument has been considered. To the extent Applicant’s argument is directed to the previous reliance on Chudak’s quantum-generated biases h and coupling strengths J as satisfying the presently claimed prior distribution, that factual basis is not maintained in the present rejection. Although the same prior-art references are retained, the present rejection relies on different disclosures and a revised rationale in view of Applicant’s amendment. As discussed in the new ground of rejection, Chudak separately teaches a GAN generator that receives samples drawn from a prior distribution r(z), and further demonstrates an operative quantum-to-generator relationship by providing information from a quantum processor to the GAN generator. Liu is relied upon for its teaching of a QCBM that evolves a quantum state through quantum gates and learns a model probability distribution. The present rejection therefore relies on the combination of Chudak and Liu to provide Liu’s quantum-generated probability distribution as the prior distribution used by Chudak’s generator.
Applicant’s argument has been considered. To the extent Applicant’s argument is directed to the previous reliance on Donahue’s externally supplied latent variable z, that factual basis is not maintained for the amended limitation. The present rejection continues to rely on Donahue, but instead relies on Donahue’s disclosure that discriminator D is a deep multi-layer network having learned intermediate representations and hidden layers followed by nonlinear activation functions. A person of ordinary skill in the art would have understood that activations produced at an internal discriminator layer form the learned intermediate representation of the input at that layer and that the representation space defined by those activations constitutes an internal latent representation space of the discriminator. Thus, Donahue teaches or at least suggests the amended internal-latent-space limitation for the reasons set forth in the new ground of rejection.
Applicant’s hindsight argument has also been considered but does not address the rejection as presently formulated. Liu teaches that quantum circuits can represent probability distributions that may be difficult to efficiently simulate using classical techniques and further teaches advantages of QCBMs for learning implicit generative models. Chudak already teaches a hybrid quantum-classical GAN environment, demonstrates that information from a quantum processor can be provided to the GAN generator, and employs a prior distribution as the source of generator input. Accordingly, a person of ordinary skill in the art would have been motivated to employ Liu’s QCBM-generated probability distribution as the prior distribution used by Chudak’s generator in order to obtain the quantum representational and learning advantages taught by Liu. The modification uses Liu’s known quantum generative technique in Chudak’s known GAN architecture according to their established functions and does not depend upon Applicant’s disclosure for the reason to combine.
Therefore, upon further consideration, new ground(s) of rejection has been raised (See Below).
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 1-4, 6-9, 12-20, 23-26 are rejected under 35 U.S.C. 103 as being unpatentable over Chudak et.al (US 20200193272 A1), further in view of Donahue et.al (NPL: Adversarial Feature Learning), further in view of Liu et.al (NPL: Differentiable Learning of Quantum Circuit Born Machine)
Chudak teaches or at least suggests the 1st limitation “a quantum computer comprising a plurality of qubits” (paragraph 61 “The quantum processor can comprise a number of qubits”, and paragraph 124 “Analog computer 350 may include an analog processor, such as quantum processor 340.” Chudak discloses systems, devices, and methods for simulating and post-processing samples generated by a hybrid computing system comprising a quantum computer and a digital computer. Within the disclosure, Chudak discloses an analog quantum computer comprising of a quantum processor which can comprise a number of qubits.)
Chudak teaches or at least suggests the 2nd limitation “a classical computer including a processor, a non-transitory computer-readable medium, and computer instructions stored in the non-transitory computer-readable medium” (paragraph 28 “A processor-based system to computationally efficiently producing sample sets, may be summarized as including: at least one processor; at least one nontransitory processor-readable medium communicatively coupled to the at least one processor and which stores processor executable instructions which, when executed by the at least one processor,” Chudak discloses the system is a processor-based system comprising of a processor, at least one non-transitory processor-readable medium in which computer instructions is stored within.)
Chudak teaches or at least suggests a part of the 3rd limitation “a generator and a discriminator operatively coupled to each other to function as a generative adversarial network (GAN) with neural network architectures for a given dataset, ...” (paragraph 52 “GANs can be useful for approximate model estimations. A GAN can include a generator and a discriminator, both of which can be multilayer perceptrons. In a typical GAN, the generator generates samples from a noise prior distribution that is defined on input noise variables, and the discriminator is trained to determine the probability of whether a sample is from the generator or from a target distribution” Chudak discloses utilizing a GAN model within the system, which comprise of the neural network architecture of a generator and a discriminator.)
Chudak teaches the 5th limitation “the computer instructions, when executed by the processor, perform a method for generating, on the hybrid quantum-classical computer system, a dataset having a plurality of datapoints, the method comprising” (paragraph 62 “FIG. 1 is a flowchart illustrating a method 100 for training an example GAN using samples generated by a quantum processor ... Method 100 can be performed by, for example, a hybrid computing system including a digital computer and a quantum processor in response to instructions or a program submitted by a user.” Chudak discloses program instructions executed by the processor as recited above to perform the method for training an example GAN using samples generated by a quantum processor, wherein the method can be performed by a hybrid computing system including a digital computer and a quantum processor.)
Chudak teaches or at least suggest the 6th limitation “initializing the sequence of instructions of the quantum component” (paragraph 123 “In some implementations, system memory 320 may store processor- or computer-readable calculation instructions to perform pre-processing, co-processing, and post-processing to analog computer 350”. Chudak discloses instructions to perform the process of the analog computer that comprises of the quantum processor, which includes the pre-processing process instruction, thereby suggesting an initializing process of the instructions of the quantum component within the claim.)
Chudak teaches or at least suggests the 7th limitation “initializing the generator and the discriminator of the (GAN)” (paragraph 95 “At 202, a generator parameter θ and a discriminator parameter ϕ are each initialized”. Chudak discloses initializing a parameter of the generator and the discriminator of the GAN network model.)
Chudak teaches or at least suggests the 8th limitation “training the GAN using prior distribution as an input to the generator of the GAN, wherein the training is performed iteratively in a first phase and a second phase” (paragraph 17 “initializing a generator parameter θ; initializing a discriminator parameter ϕ); drawing a noise sample zk from a noise prior distribution r(z); for each respective noise sample zk drawn from the noise prior distribution r(z), drawing a generated sample x(m|k) from a generator”, and paragraph 52 “A GAN can include a generator and a discriminator, both of which can be multilayer perceptrons. In a typical GAN, the generator generates samples from a noise prior distribution that is defined on input noise variables”, and paragraph 114 “Some of the exemplary acts or operations of the above described method(s), process(es), or technique(s) are performed iteratively.” Chudak discloses drawing a noise sample z_k from the noise prior distribution r(z) and using the noise sample as input to the generator to generate a sample, thereby teaching training the GAN using the prior distribution as input to the generator. Chudak further teaches that the training operations are performed iteratively.)
Chudak teaches the 9th limitation “wherein, in the first phase, the generator is not updated and the discriminator is updated” (paragraph 111 “The gradient can be updated with a respective step size γϕ for the discriminator parameter ϕ.” Chudak discloses the updating of the gradient for the discriminator parameter at the discriminator, such that a person ordinary skilled in the art would recognize that during back-propagation, a forward/backward pass may be performed on the discriminator such that the gradient is only computed at the discriminator to update the parameter of the discriminator, wherein the update is only performed at the discriminator.)
Chudak teaches the 10th limitation “wherein, in the second phase, the discriminator is not updated and the generator is updated” (paragraph 108 “The gradient can be updated with a respective step size γθ for the generator parameter θ” Chudak discloses the updating of the gradient for the generator parameter at the generator, such that a person ordinary skilled in the art would recognize that during back-propagation, a forward/backward pass may be performed on the generator such that the gradient is only computed at the generator to update the parameter of the generator, wherein the update is only performed at the generator. In other words, the forward/backward pass may be performed sequentially at each generator/discriminator model as each model get its own forward/backward and update step.)
Chudak does not teach a part of the 3rd limitation “the discriminator comprising an internal latent space defined by activations of a layer of the discriminator”. However, Donahue teaches or at least suggests this (Page 6 section 4-4.1 “In all experiments, each module D, G, and E is a parametric deep (multi-layer) network. The BiGAN discriminator D(x,z) takes data x as its initial input, and at each linear layer thereafter,… we considered alternative approaches to learning feature representations using different GAN variants … The discriminator D in a standard GAN takes data samples x ∼ pX as input, making its learned intermediate representations natural candidates as feature representations for related tasks”, and page 17 section C.1 “D, G, and E each consist of two hidden layers … The first hidden layer is followed by a non-linearity; the second is followed by (parameter-free) batch normalization … and a non-linearity … In D and E, a leaky ReLU … non-linearity with a “leak” of 0.2 is used”. Donahue discloses that discriminator D is a deep multi-layer network having learned intermediate representations, and further discloses that the hidden layers of D are followed by nonlinear activation functions, including a leaky ReLU. A person of ordinary skill in the art would have understood that the activations produced at an internal discriminator layer form the learned intermediate representation of the input at that layer, and that the representation space of the input defined by those activations constitutes an internal latent representation space of the discriminator. Thus, Donahue teaches or at least suggests the discriminator comprising an internal latent space defined by activations of a layer of the discriminator, as claimed.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of systems, devices, and methods for simulating and post-processing samples generated by a hybrid computing system comprising a quantum computer and a digital computer by Chudak with the teaching of the multi-layer discriminator with activation function at its layer by Donahue. The motivation to do so is disclosed in Donahue’s disclosure (page 6-7, section 4.1 “we considered alternative approaches to learning feature representations using different GAN variants … The discriminator D in a standard GAN takes data samples x ∼ pX as input, making its learned intermediate representations natural candidates as feature representations for related tasks … This alternative is appealing as it requires no additional machinery” Donahue discloses that the learned intermediate representations of a discriminator are natural candidates for useful feature representations for related tasks and further teaches that using such discriminator representations is appealing because it requires no additional machinery. Accordingly, a person of ordinary skill in the art would have been motivated to employ Donahue’s hidden-layer discriminator architecture and nonlinear activations in Chudak’s GAN to obtain useful learned internal representations while retaining the conventional discriminator function of the GAN, with a reasonable expectation of success because both references employ conventional multi-layer GAN discriminator architectures.)
Chudak/Donahue does not teach the 4th limitation “a quantum component, operatively coupled to the generator to provide a prior distribution to the generator, which accepts a sequence of instructions to evolve a quantum state based on a series of quantum gates”. However, Chudak in view of Liu teaches or at least suggests this limitation (Page 2 section II column 1 “Given a dataset D = {x} containing independent and identically distributed (i.i.d.) samples from a target distribution π(x), we set up a QCBM to generate samples close to the unknown target distribution. As shown in Fig. 1, the QCBM takes the product state |0i as an input and evolves it to a final state |ψθi by a sequence of unitary gates … The goal of the training is to let the model probability distribution pθ approach to π … We employ a classical-quantum hybrid feedback loop as the training strategy”, and page 3 section II column 1 “This data-driven quantum circuit architecture design scheme respects the information content of the classical dataset easier and may alleviate issues of vanishing gradients for large-scale applications”. Liu discloses a quantum circuit Born machine (QCBM) that represents and learns a model probability distribution using a quantum state, wherein the quantum state evolved through a sequence of unitary quantum gates. Thus, Liu teaches a quantum component capable of generating a quantum-based probability distribution and of evolving a quantum state based on a series of quantum gates. Chudak further demonstrates that a quantum processor may be operatively coupled to a GAN generator by providing quantum-processor-derived information as input to the generator (paragraph 96). Chudak additionally teaches that the generator operates using a noise prior distribution. Accordingly, a person of ordinary skill in the art would have understood that Liu’s QCBM-generated probability distribution could be provided through the quantum-to-generator coupling framework as taught by Chudak and used as the prior distribution for Chudak’s generator, thereby teaching or at least suggesting the claimed quantum component operatively coupled to the generator to provide a prior distribution to the generator.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of systems, devices, and methods for simulating and post-processing samples generated by a hybrid computing system comprising a quantum computer and a digital computer by Chudak, and the teaching of the multi-layer discriminator with activation function at its layer by Donahue with the teaching of a QCBM for generating quantum probability distribution through quantum-state evolution by Liu. The motivation to do so is referred to in Liu’s disclosure (page 1 section 1 “Moreover, computational complexity considerations on quantum sampling problems suggest that a quantum circuit can produce probability distribution that is #P-hard … which is infeasible to simulate efficiently using classical algorithms”, and Page 3 section II-B column 2 “the unbiased gradient estimator of the QCBM takes advantage of the known structure of the unitary evolution and the MMD loss (see Appendix A), despite that the probability of the outcome is unknown. In this sense, quantum circuits exhibit a clear quantum advantage over classical neural nets since they fill the gap of differentiable learning of implicit generative models of discrete data”. Liu discloses that quantum circuits can represent probability distributions that may be difficult to efficiently simulate using classical techniques and further teaches advantages of QCBMs for learning implicit generative models. Accordingly, a person of ordinary skill in the art would have been motivated to employ Liu’s QCBM-generated probability distribution as the prior distribution provided to Chudak’s GAN generator in order to obtain the quantum representational and learning advantages taught by Liu. Chudak already demonstrates that a quantum processor may be operatively coupled to and communicate with the GAN generator, such that applying Liu’s quantum-generated distribution as Chudak’s generator prior would have been a predictable use of a known quantum generative technique in Chudak’s known hybrid GAN architecture, with a reasonable expectation of success.)
Regarding claim 2 depends on claim 1, thus the rejection of claim 1 is incorporated
Liu teaches the limitation “The system of claim 1, wherein training the GAN further comprises training the quantum component” (Page 3 section II-B column 1 “Viewing the QCBM as an implicit generative model ... we train it by employing the kernel two-sample test ... We refer the following loss function as the squared maximum mean discrepancy (MMD)” Liu discloses the training of the QCBM by viewing it as a generative model and apply a loss function to train the model.)
Regarding claim 3 depends on claim 2, thus the rejection of claim 2 is incorporated
Liu teaches the limitation “The system of claim 2, wherein training the quantum component comprises training the quantum component based on a cost function” (Page 3 section II-B column 1 “Viewing the QCBM as an implicit generative model ... we train it by employing the kernel two-sample test ... We refer the following loss function as the squared maximum mean discrepancy (MMD)” Liu discloses the training of the QCBM by viewing it as a generative model and apply a loss function to train the model, wherein the loss function suggests a cost function as understood by a person ordinary skilled in the art.)
Regarding claim 4 depends on claim 2, thus the rejection of claim 2 is incorporated.
Donahue teaches this limitation (Page 2 section 1 “Hence, we propose a novel unsupervised feature learning framework, Bidirectional Generative Adversarial Networks (BiGAN)... The BiGAN discriminator D discriminates not only in data space (x versus G(z)), but jointly in data and latent space”, and Page 3 section 3 “The discriminator is also modified to take input from the latent space,”. Donahue discloses the discriminator of the BiGANs take input from the latent space to discriminate, suggesting that the discriminator within the network architecture is associated with the latent space containing data samples, such that a person ordinary skilled in the art may incorporate the BiGANs during training of the quantum computer disclosed above to incorporate the inverse mapping of the discriminator associated with the latent space.)
Regarding claim 6 depends on claim 1, thus the rejection of claim 1 is incorporated
Liu teaches the limitation “The system of claim 1, wherein initializing the sequence of instructions of the quantum component comprises evolving a quantum state of the quantum component, to produce an evolved quantum state, such that measurements of the evolved quantum state produce output samples of the prior distribution according to a desired probability distribution.” (Page 2 section II column 1 “As shown in Fig. 1, the QCBM takes the product state |0i as an input and evolves it to a final state |ψθi by a sequence of unitary gates ... Then we can measure this output state on computation basis to obtain a sample of bits x ∼ pθ(x) ... The goal of the training is to let the model probability distribution pθ approach to π.” Liu discloses evolving an initial quantum state of the QCBM through a sequence of unitary quantum gates to produce an evolved quantum state, and measuring the evolved state in the computational basis to obtain outputs according to the model probability distribution pθ. Liu further teaches training the QCBM such that pθ approaches the desired probability distribution π. As discussed with respect to claim 1, pθ is employed as the prior distribution provided to the generator in the combined system. Thus, Liu teaches or at least suggests evolving the quantum state such that measurements of the evolved quantum state produce outputs from the prior distribution according to a desired probability distribution, as claimed.)
Regarding claim 7 depends on claim 6, thus the rejection of claim 6 is incorporated
Chudak teaches the limitation “wherein the desired probability distribution is uniform over a selected range.” (paragraph 64 “At 104, a noise sample zk is drawn from a noise prior distribution r(z). The noise prior distribution can be a fixed distribution. For example, the noise prior distribution can be a uniform distribution ... (i.e., z ∈ [0,1] or z ∈ E (0,1)).” Chudak discloses the analog computer can be operated to provide samples from a probability distribution, wherein the probability distribution in which the samples are drawn from can be a uniform distribution over a range. A person ordinary skilled in the art would have been able to configure such that the probability distribution approach to π as desired as disclosed above by Liu is uniform over a selected range.)
Regarding claim 8 depends on claim 1, thus the rejection of claim 1 is incorporated
Liu teaches the limitation “wherein the quantum component is a quantum circuit born machine (QCBM).” (Page 3 section III column 2 “We carry out numerical experiments by simulating the learning of QCBM on a classical computer” Liu discloses the embodiment is carried out by simulating the QCBM on a classical computer, which suggest that it may be incorporated as part of the analog computer with the quantum processor by Chudak.)
Regarding claim 9 depends on claim 1, thus the rejection of claim 1 is incorporated
Liu teaches the limitation “The system of claim 1, further comprising measuring the quantum component using a multi-basis method.” (Page 2 section II column 1 “...Then we can measure this output state on computation basis to obtain a sample of bits”, and Page 3 section II-B column 2 “In order to estimate the gradient [Eq. (2)] on an actual quantum circuit, one can repeatedly send rotation and entangle pulses to the device according to the circuit parameters θ (±), and then perform projective measurements on the computational basis to collect binary sample”. Liu discloses perform the measurement of the output state of the QCBM on computation basis, wherein the computation basis suggests the multi-basis method as understood by a person ordinary skilled in the art, as the computational basis refers to a fundamental set of orthonormal states used to represent quantum information, often represented as |0⟩ and |1⟩ for a single qubit and are the building blocks for describing any quantum state within a system.)
Regarding claim 12 depends on claim 1, thus the rejection of claim 1 is incorporated.
Donahue teaches the limitation “The system of claim 1, wherein the dataset includes higher-resolution handwritten digits” (Page 7 figure 2 “Qualitative results for permutation-invariant MNIST BiGAN training, including generator samples G(z), real data x, and corresponding reconstructions G(E(x))” Donahue discloses using permutation-invariant MNIST test data set as illustrated in figure 2 which comprise of a higher-resolution handwritten digits after the training of the BiGAN model.)
Regarding claim 13 depends on claim 1, thus the rejection of claim 1 is incorporated.
Donahue teaches the limitation “The system of claim 1, wherein the dataset includes monochrome images and color images” (Page 8 figure 4 “Figure 4: Qualitative results for ImageNet BiGAN training, including generator samples G(z), real data x, and corresponding reconstructions G(E(x)).” Donahue discloses using ImageNet data et as illustrated in figure 4 which comprise of monochrome images and color images for the training of the BiGAN model.)
Regarding claim 14 depends on claim 1, thus the rejection of claim 1 is incorporated.
Donahue teaches the limitation “The system of claim 1, wherein the dataset includes video frames” (Page 7-8 figure 4 “Next, we present results from training BiGANs on ImageNet LSVRC (Russakovsky et al., 2015), a large-scale database of natural images.” Donahue discloses results from training the BiGAN on ImageNet LSVRC, which is a large scale database of natural images, wherein a person ordinary skilled in the art would recognize that this database of images comprise of video frame such that the training result in Figure 4 may comprise of video frame as well.)
Regarding claim 15 depends on claim 1, thus the rejection of claim 1 is incorporated
Liu teaches the limitation “The system of claim 1, wherein the plurality of qubits includes at least 8-qubits.” (Page 3 section III-A column 1 “We first train a QCBM on the Bars-and-Stripes dataset ... We model the pixels with a quantum circuit of 9 qubits.” Liu discloses training a QCBM and using a quantum circuit of 9 qubits.)
Regarding claim 16, the applicant is further directed to the rejection of claim 1 above, because claim 16 comprise of similar limitations to claim 1, thus they are rejected based on the same rationale.
Regarding claim 17, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 2 above, because claim 17 comprise of similar limitations to claim 2, thus they are rejected based on the same rationale.
Regarding claim 18, depends on claim 17, thus the rejection of claim 17 is incorporated. The applicant is further directed to the rejection of claim 3 above, because claim 18 comprise of similar limitations to claim 3, thus they are rejected based on the same rationale.
Regarding claim 19, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 8 above, because claim 19 comprise of similar limitations to claim 8, thus they are rejected based on the same rationale.
Regarding claim 20, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 9 above, because claim 9 comprise of similar limitations to claim 9, thus they are rejected based on the same rationale.
Regarding claim 23, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 12 above, because claim 12 comprise of similar limitations to claim 23, thus they are rejected based on the same rationale.
Regarding claim 24, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 13 above, because claim 13 comprise of similar limitations to claim 24, thus they are rejected based on the same rationale.
Regarding claim 25, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 14 above, because claim 14 comprise of similar limitations to claim 25, thus they are rejected based on the same rationale.
Regarding claim 26, depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 7 above, because claim 7 comprise of similar limitations to claim 26, thus they are rejected based on the same rationale.
Claims 5, 11, are rejected under 35 U.S.C. 103 as being unpatentable over Chudak et.al (US 20200193272 A1), further in view of Donahue et.al (NPL: Adversarial Feature Learning), further in view of Liu et.al (NPL: Differentiable Learning of Quantum Circuit Born Machine), further in view of Arici et.al (NPL: Associative Adversarial Networks)
Regarding claim 5 depends on claim 1, thus the rejection of claim 1 is incorporated
Chudak/Donahue/Liu does not teach the limitation “The system of claim 1, wherein a dimension of the latent space equals a dimension of the input of the generator”. However, Arici teaches this limitation (page 4 section 5 “The discriminator model D tries to discriminate between the real and fake data. In doing so, D learns features that can explain factors of variation in data and uses these features to achieve its classification goal. On the other hand G tries to map a low-dimensional input to data. Contrary to common approaches, instead of representing the G’s input as a flat space, we think of it as an intermediate but higher-level representation corresponding to one of the intermediate layers of D … Therefore we connect D and G through a high-level feature space. Let F(x) denote an intermediate layer activations of D, while C(y) denotes the operations in the remaining layers in D. Then, D(x) = C(F(x))”, and page 5 section 6-6.1 “The negative samples created for RBM’s contrastive divergence learning are used as inputs for the G … The last convolutional layer’s outputs are reshaped into a one-dimensional representation f which is fully connected to a layer that has the same dimension as the RBM’s visible layer”. Arici discloses that an intermediate layer of the discriminator produces a high-level feature representation and is connected to an RBM visible layer having the same dimension. Arici further discloses that the RBM learns a probability distribution corresponding to the discriminator-derived features and generates samples that are provided as inputs to the generator. Because the RBM visible layer has the same dimension as the discriminator’s intermediate feature representation, and because the RBM-generated samples are supplied to the generator, Arici discloses that the discriminator’s selected internal feature layer and the generator input have corresponding dimensions. Accordingly, Arici expressly discloses the claimed dimensional arrangement. A person of ordinary skill in the art would have understood that this same arrangement could be incorporated into the Chudak/Donahue/Liu combination by configuring Donahue’s internal discriminator latent space and Chudak’s generator input with equal dimensions, while using Liu’s QCBM to provide the corresponding learned prior distribution, thereby teaching or at least suggesting the claimed limitation.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of Chudak/Donahue/Liu with the teaching of configuring the discriminator’s internal feature representation and the generator input with corresponding dimensionality through a learned feature-space distribution by Arici.The motivation to do so is referred to in Arici’s disclosure (page 1-2 section 1 “In this paper, we argue that using noise for the generator contributes to the difficulty of training GANs. The generator is assigned the task of learning the mapping from the input signal z, which is a uniformly distributed noise to data space … a strong evidence is presented for better disentangling of the underlying factors of variation in data by higher-level representation spaces, and this might explain why a uniformly distributed random noise as input to G works. However, learning the mapping from a flat representation space to data space is difficult … By using an associative memory on z and generating samples from this memory as an input to G, our goal is to alleviate G’s learning task” Arici discloses that replacing conventional noise with a learned higher-level representation of the discriminator’s internal features alleviates the generator’s learning task. Arici further teaches an arrangement in which the discriminator’s selected internal feature representation is connected to the visible layer of the RBM, and the RBM generates samples that are provided to the generator. A person of ordinary skill in the art would therefore have been motivated to configure the discriminator’s internal latent space and the generator input with equal dimensions, corresponding to the equal-dimensional connection between the discriminator feature representation, the RBM visible layer, and the generator input disclosed by Arici. This arrangement would allow the learned feature distribution to be transferred directly through the RBM or QCBM without an additional dimensionality-changing transformation, thereby avoiding unnecessary architectural complexity and simplifying the generator’s learning task. Accordingly, the Chudak/Donahue/Liu combination could incorporate Arici’s feature-space arrangement by using Liu’s QCBM in place of Arici’s classical RBM to provide the generator with a corresponding learned feature distribution.)
Regarding claim 11 depends on claim 5, thus the rejection of claim 5 is incorporated.
Liu teaches the limitation “The system of claim 5, wherein training the quantum component further comprises calculating a loss function for the quantum component based on the prior distribution” (Page 3 section II-B column 1 “Viewing the QCBM as an implicit generative model ... we train it by employing the kernel two-sample test ... We refer the following loss function as the squared maximum mean discrepancy (MMD)” Liu discloses training the QCBM by calculating a squared maximum mean discrepancy (MMD) loss that measures the difference between the QCBM model probability distribution pθ and the desired target distribution π, and using gradients of the loss to train the quantum circuit. Because the QCBM model probability distribution pθ corresponds to the prior distribution provided to the generator in the combined system, Liu teaches or at least suggests calculating a loss function for the quantum component based on the prior distribution, as claimed.)
Claims 10, 21, 27 are rejected under 35 U.S.C. 103 as being unpatentable over Chudak et.al (US 20200193272 A1), further in view of Donahue et.al (NPL: Adversarial Feature Learning), further in view of Liu et.al (NPL: Differentiable Learning of Quantum Circuit Born Machine), further in view of Shehab et.al (US 20200372390 A1)
Regarding claim 10 depends on claim 1, thus the rejection of claim 1 is incorporated
Chudak/Donahue/Liu does not teach the limitation “The system of claim 1, wherein the quantum component comprises a trapped-ion quantum device”. However, Shehab teaches this limitation (paragraph 0006 “For solving some optimization problems, a NISQ device having shallow circuits (with small number of gate operations to be executed in time-sequence) can be used in combination with a classical computer (referred to as a hybrid quantum-classical computing system... The classical computer (also referred to as a “classical optimizer”) instructs a controller to prepare the NISQ device (also referred to as a “quantum processor”) in an N-qubit state, execute quantum gate operations, and measure an outcome of the quantum processor.”, and paragraph 23 “The quantum processor includes trapped ions that are coupled with various hardware, including lasers to manipulate internal hyperfine states (qubit states) of the trapped ions and an acousto-optic modulator to read-out the internal hyperfine states (qubit states) of the trapped ions.” Shehab discloses the incorporation of an NISQ device for many problems, wherein the device can be referred to as a quantum processor. The quantum processor includes trapped ions that are coupled with various hardware.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of Chudak/Donahue/Liu with the teaching of Noise reduced circuits for trapped-ion quantum computers by Shehab. The motivation to do so is referred to in Shehab’s disclosure (paragraph 0006 “For solving some optimization problems, a NISQ device having shallow circuits ... The classical computer (also referred to as a “classical optimizer”) instructs a controller to prepare the NISQ device (also referred to as a “quantum processor”) in an N-qubit state, execute quantum gate operations, and measure an outcome of the quantum processor.” Shehab discloses the NISQ device may be referred to as a quantum processor, wherein Chudak also discloses a quantum processor of the analog computer. Therefore, a person ordinary skilled in the art would have been able to recognize that the NISQ device as referred to in Shehab is similar to the device in Chudak and the embodiment of the NISQ device by Shehab may be applied to the quantum processor as disclosed in Chudak and the teaching combination. Shehab further discloses improvement of using a hybrid quantum-classical computing system at paragraph 44 “This hybrid quantum-classical computing system has at least the following advantages. First, an initial guess is derived from a classical computer, and thus the initial guess does not need to be constructed in a quantum processor that may not be reliable due to inherent and unwanted noise in the system. Second, a quantum processor performs a small-sized (e.g., between a hundred qubits an a few thousand qubits) but accelerated operation (that can be performed using a small number of quantum logic gates) between an input of a guess from the classical computer and a measurement of a resulting state, and thus a NISQ device can execute the operation without accumulating errors. Thus, the hybrid quantum-classical computing system may allow challenging problems to be solved, such as small but challenging combinatorial optimization problems, which are not practically feasible on classical computers, or suggest ways to speed up the computation with respect to the results that would be achieved using the best known classical algorithm”, thus the teaching combination may further improve upon the teaching by Shehab.)
Regarding claim 21 depends on claim 16, thus the rejection of claim 16 is incorporated. The applicant is further directed to the rejection of claim 10 above, because claim 21 comprise of similar limitations to claim 10, thus they are rejected based on the same rationale.
Regarding claim 27 depends on claim 16, thus the rejection of claim 16 is incorporated
Chudak/Donahue/Liu does not teach the limitation “The method of claim 16, wherein the quantum component is a noisy intermediate-scale (NISQ) device” However, Shehab teaches this limitation (paragraph 0006 “For solving some optimization problems, a NISQ device having shallow circuits (with small number of gate operations to be executed in time-sequence) can be used in combination with a classical computer (referred to as a hybrid quantum-classical computing system... The classical computer (also referred to as a “classical optimizer”) instructs a controller to prepare the NISQ device (also referred to as a “quantum processor”) in an N-qubit state, execute quantum gate operations, and measure an outcome of the quantum processor.”, Shehab discloses the NISQ device for many problems, wherein the device can be referred to as a quantum processor.)
The motivation to combine the teaching combination with the teaching by Shehab is similar to the motivation as recited in claim 10.
Claim 22 is rejected under 35 U.S.C. 103 as being unpatentable over Chudak et.al (US 20200193272 A1), further in view of Donahue et.al (NPL: Adversarial Feature Learning), further in view of Liu et.al (NPL: Differentiable Learning of Quantum Circuit Born Machine), further in view of Dallaire-Demers et.al (NPL: Quantum generative adversarial networks)
Regarding claim 22 depends on claim 16, thus the rejection of claim 16 is incorporated
Chudak/Donahue/Liu does not teach the limitation “The method of claim 16, wherein a QC – AAN framework is used for the quantum component”. However, Dallaire-Demers teaches this limitation (Page 1 “In this work and a companion paper, we extend adversarial training to the quantum domain and show how to construct generative adversarial networks using quantum circuits.”, Dallaire-Demers discloses an embodiment of constructing a generative adversarial network using quantum circuits, which suggest a QC – AAN framework.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of Chudak/Donahue/Liu with the teaching of quantum generative adversarial networks by Dallaire-Demers. The motivation to do so is referred to in Dallaire-Demers’s disclosure (Page 8 “It is expected that QuGANs will have a more versatile representation power than their classical counterpart. For example, one can speculate that a large enough QuGAN could learn to generate encrypted data labeled by RSA public encryption key ... In this work, we have explored the practical issues of QuGANs, namely, explicit quantum circuits for the generator and discriminator, as well as quantum methods for computing the gradients of these circuits” Dallaire-Demers discloses how to construct a generative adversarial network using quantum circuits, as well as its improvement over the conventional GAN by employing the quantum circuit. Therefore, a person ordinary skilled in the art may incorporate the teaching of Dallaire-Demers into the teaching combination for further improvement.)
Claims 28 is rejected under 35 U.S.C. 103 as being unpatentable over Chudak et.al (US 20200193272 A1), further in view of Donahue et.al (NPL: Adversarial Feature Learning), further in view of Liu et.al (NPL: Differentiable Learning of Quantum Circuit Born Machine), further in view of Zhang et.al (NPL: Progressive Augmentation of GANs), further in view of Anschuetz et.al (NPL: Near-Term Quantum-Classical Associative Adversarial Networks)
Regarding claim 28 depends on claim 16, thus the rejection of 16 is incorporated.
Chudak/Donahue/Liu does not teach a part of the limitation “The method of claim 16, wherein the latent space is increased in the discriminator …” However, Zhang teaches or at least suggests this (page 5 section 3.2 “Progressive augmentation. With the aim of maximally reusing existing GAN architectures we propose two augmentation options ... The second option is feature space augmentation, where s is concatenated with the learned feature representations of x attained at intermediate hidden layers … For both cases, the way to concatenate s with x or its feature maps is identical. Each entry sl creates one augmentation channel, which is replicated to match the spatial dimension of x or its feature maps … channel size being increased by l … When a new augmentation level is reached, one extra input channel of the filter is instantiated to process the bit l + 1”. Zhang discloses progressively augmenting the discriminator’s internal feature representations at intermediate hidden layers by concatenating additional feature channels, wherein the corresponding channel dimension is increased to accommodate the augmented representation. A person of ordinary skill in the art would have understood that the increased number of feature channels enlarges the dimensionality of the internal feature representation processed by the discriminator. Accordingly, applying Zhang’s feature-space augmentation to the internal latent representation of Donahue’s discriminator would increase the dimensionality of the discriminator’s internal latent space, thereby teaching or at least suggesting the claimed increase in the latent space of the discriminator.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of Chudak/Donahue/Liu with the teaching of progressively augmenting the discriminator’s internal feature representations at intermediate hidden layers by increasing the corresponding feature-channel dimension by Zhang. The motivation to do so is referred to in Zhang’s disclosure (page 5 section 3.2 “These two ways of augmentation are beneficial as they make the checksum computation more challenging for the discriminator, i.e., making the discriminator unaware about the need of separating x and s from the concatenated input”, and page 1 section 1 “The key idea is to progressively augment the input of the discriminator network or its intermediate feature layers with auxiliary random bits in order to gradually increase the discrimination task difficulty (see Fig. 1). In doing so, the discriminator can be prevented from becoming over-confident, enabling continuous learning of the generator”. Zhang discloses that progressively augmenting the discriminator’s intermediate feature space increases the difficulty of the discriminator’s task and helps prevent the discriminator from becoming overconfident, thereby enabling continued learning of the generator. Accordingly, a person of ordinary skill in the art would have been motivated to apply Zhang’s feature-space augmentation to the internal latent representation of Donahue’s discriminator in the Chudak/Donahue/Liu combination to increase the discriminator’s internal feature-space dimensionality and maintain effective adversarial training, thereby obtaining the training benefits taught by Zhang.)
Chudak/Donahue/Liu/Zhang does not teach a part of the limitation “… wherein the method further comprises training quantum component to output the prior distribution based on activations of the discriminator” However, Anschuetz teaches or at least suggests this (page 1 section abstract “a small auxiliary quantum Boltzmann machine that is simultaneously trained on an intermediate layer of the discriminator of the generative network”, and page 4 section II-C “The RBM is simultaneously trained with the GAN to learn the distribution of a layer of the discriminator of the GAN. The distribution approximated by the RBM then acts as the latent distribution for the generator of the GAN (instead of noise). A QAAN has an identical architecture, except that the RBM is replaced with a QBM”, page 4 section III-B “We now quantize AANs by transforming the associated RBM into a QBM, and call the resulting architecture a quantum-classical associative adversarial network (QAAN). Our implementation otherwise exactly follows that of our AAN”. Anschuetz discloses a quantum-classical associative adversarial network (QAAN) comprising a quantum Boltzmann machine (QBM) that is simultaneously trained on an intermediate layer of the discriminator. Anschuetz further discloses that the quantum model learns a probability distribution corresponding to the discriminator’s internal feature representation, wherein the learned distribution serves as the latent distribution for the generator, replacing conventional noise. Accordingly, Anschuetz teaches training a quantum component based on activations of the discriminator to learn and provide a prior distribution to the generator. A person of ordinary skill in the art would have understood that Anschuetz’s discriminator-informed quantum training arrangement could be incorporated into the Chudak/Donahue/Liu/Zhang combination by training Liu’s quantum component based on activations of the modified discriminator layer to output the corresponding prior distribution for Chudak’s generator, thereby teaching or at least suggesting the claimed limitation.)
Before the effective filing date, it would have been obvious to a person ordinary skilled in the art to combine the teaching of Chudak/Donahue/Liu/Zhang with the teaching of training a quantum Boltzmann machine based on an intermediate discriminator layer to learn a prior distribution by Anschuetz. The motivation to do so is referred to in Anschuetz’s disclosure (page 1 section 1 “One of the major reasons GANs are difficult to train is that the generator and discriminator learn at different rates generically. One approach towards combating this issue is through the use of associative adversarial networks … In this architecture, a Boltzmann machine acts as an associative memory that learns the high-level feature distribution of a layer of the discriminator. The generator then draws samples from the distribution ap proximated by the Boltzmann machine. The associative memory also adds expresivity to the network by providing the generator with inputs drawn from a more meaningful, data-specific distribution, rather than a uniform or Gaussian distribution as is the case for standard GANs … Motivated by the observed improvement in per formance of QBMs over RBMs … we propose a method of implementing hybrid quantum-classical AANs (QAANs), where the associative memory is instead provided by a QBM.” Anschuetz discloses that training a quantum Boltzmann machine based on an intermediate discriminator layer enables the quantum model to learn a more meaningful, data-specific prior distribution for the generator, thereby improving network expressivity and addressing difficulties associated with different learning rates of the generator and discriminator. Accordingly, a person of ordinary skill in the art would have been motivated to incorporate Anschuetz’s discriminator-informed quantum training technique into the Chudak/Donahue/Liu/Zhang combination by training Liu’s quantum component based on activations of the modified discriminator’s internal layer to learn and output a prior distribution corresponding to the enlarged feature representation for Chudak’s generator, thereby obtaining the data-specific prior distribution and improved expressivity taught by Anschuetz.)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DUY TU DIEP whose telephone number is (703)756-1738. The examiner can normally be reached M-F 8-4:30.
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.
/DUY T DIEP/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights