DETAILED ACTION
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 .
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.
Claim(s) 1-2, 7, 9-10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shiv Shankar (Hereinafter Shankar) (“Neural Variational Boson Sampling”, 09/01/2022) in view of Nathan Killoran et al. (Hereinafter Killoran) (“Continuous-variable quantum neural networks”, 10/31/2019).
Regarding claim 1,
Shankar teaches;
A method for generating a dataset, the method comprising: controlling a boson sampler to produce one or more integer sequences,
[pg. 2]
PNG
media_image1.png
59
375
media_image1.png
Greyscale
…
PNG
media_image2.png
89
392
media_image2.png
Greyscale
NOTE: Teaches the Boson sampler being controlled to produce one or more integer sequences (the output patterns of the boson sampler comprise integer sequences, as pictured by n-hat).
each of the one or more integer sequences determined from a measurement outcome of
[pg. 2]
PNG
media_image2.png
89
392
media_image2.png
Greyscale
NOTE: Teaches each of the one or more integer sequences determined from a measurement outcome of the boson sampler (n-hat represents a measured output pattern of photons of the boson sampler).
determining, from the one or more integer sequences, one or more latent vectors;
([pg. 5] The NBS model sample vectors from a 5-dimensional latent space which then constructs the images through a classical conditional generator... Z is a latent variable... z and y are used to denote samples of the variables Z and Y respectively... The distribution of the latent variable z is given by the output distribution of the Boson sampler)
NOTE: z is a latent vector (a vector sampled from latent variable Z) determined from the one or more integer sequences (the aforementioned output distributions of the Boson sampler, n-hat, is instantiated as the latent vector z).
providing the determined one or more latent vectors to a trained {model}
[pg. 5]
PNG
media_image3.png
35
375
media_image3.png
Greyscale
…
PNG
media_image4.png
52
377
media_image4.png
Greyscale
NOTE: Teaches a generative model (the conditional generator) which is trained (has learned/trained parameters).
([pg. 5] The NBS model sample vectors from a 5-dimensional latent space which then constructs the images through a classical conditional generator. The complete generative model is depicted in Figure 5. In the figure, Y is a random variable that denotes the training images, Z is a latent variable, and Pn is the distribution of Z parameterized by D. The parameters Ѱ are the parameters of the conditional generator. z and y are used to denote samples of the variables Z and Y respectively.)
NOTE: Teaches providing the determined one or more latent vectors (z) to a trained {model} (the aforementioned classical conditioner) configured to convert the one or more latent vectors to a generated dataset (the classical conditional generator constructs images, where a generated collection of images can be considered a generated dataset) and outputting the generated dataset (outputs the generated images).
Shankar fails to teach but Killoran teaches;
one or more photodetectors
artificial neural network (ANN)
([pg. 9] We will now overview how specific models of photonic quantum computing can be realized using CV quantum neural networks. First, Gaussian boson sampling (GBS) [106] is a model of photonic quantum computing where a multimode Gaussian state is prepared and subsequently measured in the photon-number basis, a procedure which is believed to be hard to simulate classically. GBS includes conventional Boson Sampling as a special case [107,108]. Any GBS configuration can be encoded in a CV quantum neural network by (i) turning off the non-Gaussian gates and (ii) measuring the outputs using photon detectors.)
NOTE: Discloses how Gaussian boson sampling can be realized using quantum neural networks, teaching measuring outputs with one or more photodetectors of the boson sampler (photon detectors are a type of photodetector) and an ANN (quantum neural network).
wherein the ANN comprises one or more hidden layers ([pg. 13] The input layer then feeds onto four hidden layers with fully controllable parameters) implementing nonlinear activation functions ([Abstract] nonlinear activation functions, … are enacted in the quantum network using … non-Gaussian gates … [pg. 5] The circuit structure for a single layer of a CV quantum neural network: … local non-Gaussian gates);
OBVIOUSNESS TO COMBINE SHANKAR WITH KILLORAN:
Shankar and Killoran are both analogous to each other and to the present disclosure as they both pertain to machine learning and quantum computing. Specifically, Shankar gives a boson sampling generative model use case, and Killoran gives a known quantum neural-network architecture for implementing neural network behavior on photonic hardware and reading outputs by photonic measurement.
Killoran additionally states,
([pg. 2] In this work, we show that the CV model gives a native architecture for building neural network models on quantum computers.)
NOTE: The CV model (the architecture which Killoran uses in their disclosure) provides a native architecture for building neural networks in quantum computing environments.
From this, Killoran teaches a native architecture for building a neural network in a quantum computing environment which could be used to implement the model for dataset generation as taught by Shankar (as Shankar pertains to a quantum computing environment).
Killoran additionally states;
([pg. 2] Building machine learning models with multilayer neural networks is well-motivated because of various universality theorems [54–56]. These theorems guarantee that, provided enough free parameters, feedforward neural networks can approximate any continuous function on a closed and bounded subset of Rn to an arbitrary degree of accuracy.)
NOTE: This excerpt discloses that building machine learning models using neural networks is beneficial because of various universality theorems.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to use the recognized neural-network framework as taught by Killoran to host a neural network be utilized to implement the boson sampler derived dataset generation model as taught by Shankar, to improve the universality of the model.
Regarding claim 2, Shankar teaches;
A method comprising: controlling a boson sampler to produce a set of integer sequences,
[pg. 2]
PNG
media_image1.png
59
375
media_image1.png
Greyscale
…
PNG
media_image2.png
89
392
media_image2.png
Greyscale
NOTE: Teaches the Boson sampler being controlled to produce integer sequences (the output patterns of the boson sampler comprise integer sequences, as pictured by n-hat)
([pg. 2] The BosonSampling (BS) problem refers to sampling outcomes from a linear optical network.)
NOTE: Teaches controlling the boson sampler producing a set of integer sequences ('sampling outcomes' indicates a plurality of outcomes, i.e. a set, the outcomes being the aforementioned n-hat which is an integer sequence)
determining, from the set of integer sequences, a first set of latent vectors;
([pg. 5] The NBS model sample vectors from a 5-dimensional latent space which then constructs the images through a classical conditional generator... Z is a latent variable... z and y are used to denote samples of the variables Z and Y respectively... The distribution of the latent variable z is given by the output distribution of the Boson sampler)
NOTE: z is a latent vector (a vector sampled from latent variable Z) determined from the set of integer sequences (the aforementioned output distributions of the Boson sampler, n-hat, is instantiated as the latent vector z).
[pg. 5]
PNG
media_image5.png
273
828
media_image5.png
Greyscale
NOTE: The term ‘samples’ is plural here and therefore indicates a plurality, i.e. a set of samples z. From this, the pictured samples can be considered a first set of determined latent vectors.
and using the determined first set of latent vectors, training a {model}
[pg. 5] We train the model to maximize the log-likelihood of the data with a variant of the EM algorithm [Dempster et al., 1977]. In the EM algorithm, each iteration consists of repeated application of the E-step and the M-step. In the E-step, the data log-likelihood conditioned on the observed variables is computed. On the other hand, in the M-step, the likelihood obtained in the E-step is maximized with respect to the model parameters. The Monte-Carlo Expectation Maximization (MCEM) algorithm [Wei and Tanner, 1990] is a variant of the classic EM; often used for high-dimensional data or when the integral required in the E-step is intractable… Specifically, the Q function for the MCEM algorithm in our case is given by:
PNG
media_image6.png
191
824
media_image6.png
Greyscale
NOTE: The disclosed image generation EM loop of Shankar is essentially:
Start with current parameters.
Sample a first set of latent vectors (the aforementioned first set of latent vectors) from the boson sampler.
Feed the normalized latent vectors into the Gaussian decoder to see how well they explain each image.
Update parameters.
Which teaches training a model using a determined first set of latent vectors.
([pg. 5] The NBS model sample vectors from a 5-dimensional latent space which then constructs the images through a classical conditional generator. The complete generative model is depicted in Figure 5. In the figure, Y is a random variable that denotes the training images, Z is a latent variable, and Pn is the distribution of Z parameterized by D. The parameters Ѱ are the parameters of the conditional generator. z and y are used to denote samples of the variables Z and Y respectively.)
NOTE: Teaches sampling a second set of latent vectors (samples vectors z [plural form of ‘vector’ indicates a plurality, i.e., a set] from a latent space) which are converted to a generated dataset (which then constructs images through a classical conditional generator, where a set of generated images is considered a generated dataset) using the model trained from the determined first set of latent vectors.
Shankar fails to teach but Killoran teaches;
each integer sequence [taught by Shankar] determined from a measurement outcome of photodetectors of the boson sampler;
artificial neural network (ANN)
([pg. 9] We will now overview how specific models of photonic quantum computing can be realized using CV quantum neural networks. First, Gaussian boson sampling (GBS) [106] is a model of photonic quantum computing where a multimode Gaussian state is prepared and subsequently measured in the photon-number basis, a procedure which is believed to be hard to simulate classically. GBS includes conventional Boson Sampling as a special case [107,108]. Any GBS configuration can be encoded in a CV quantum neural network by (i) turning off the non-Gaussian gates and (ii) measuring the outputs using photon detectors.)
NOTE: Discloses how Gaussian boson sampling can be realized using quantum neural networks, teaching measuring outputs (where the outputs could be the aforementioned integer sequences taught by Shankar) with one or more photodetectors of the boson sampler (photon detectors are a type of photodetector) and an ANN (quantum neural network).
wherein the ANN comprises one or more hidden layers ([pg. 13] The input layer then feeds onto four hidden layers with fully controllable parameters) implementing nonlinear activation functions ([Abstract] nonlinear activation functions, … are enacted in the quantum network using … non-Gaussian gates … [pg. 5] The circuit structure for a single layer of a CV quantum neural network: … local non-Gaussian gates);
It would be obvious for the generative model of Shankar to be implemented using a quantum neural network and architecture disclosed by Killoran, using the same obviousness rational from claim 1.
Regarding claim 7, Shankar teaches;
The method according to claim 2, wherein the set of integer sequences comprises a set of binary strings.
[pg. 4]
PNG
media_image7.png
736
468
media_image7.png
Greyscale
NOTE: Teaches the aforementioned integer sequences comprising a binary vector. A binary vector can be considered a binary string.
Regarding claim 9,
Claim 9 is a computer readable medium claim directly corresponding to claim 1, and is therefore rejected using the same reasoning.
Regarding claim 10,
Claim 10 is a computer readable medium claim directly corresponding to claim 2, and is therefore rejected using the same reasoning.
Claim(s) 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Ian J. Goodfellow et al. (hereinafter Goodfellow) (“Generative Adversarial Nets”, 06/10/2014)
Regarding claim 3, Shankar in view of Killoran teaches;
The method according to claim 2,
(Using the same reasoning from the claim 2 rejection)
Shankar teaches;
the first and second set of latent vectors
(Using the same reasoning from the claim 2 rejection)
Shankar and Killoran fail to teach but Goodfellow teaches;
wherein training the ANN to convert
([pg. 2] The adversarial modeling framework is most straightforward to apply when the models are both multilayer perceptrons.)
NOTE: Discloses that their models are ideally multilayer perceptrons, which are considered artificial neural networks.
([Abstract] We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G.)
NOTE: Teaches training the ANN (train two models, each being an ANN, as previously disclosed) to convert data to a generated dataset (the generative model G generates a dataset from input data) comprising training a generative adversarial network comprising the ANN (G) and a second ANN (D).
and wherein training the GAN comprises: training the ANN,
([pg. 2] To learn the generator’s distribution pg over data x, we define a prior on input noise variables pz(z), then represent a mapping to data space as G(z;θg), where G is a differentiable function represented by a multilayer perceptron with parameters θg... We simultaneously train G to minimize log(1 − D(G(z))))
NOTE: Teaches training the ANN (G is trained) using the determined set of noise variables pz(z) (where noise variables are similar to the latent vectors, as they are both hidden/unobserved vectors) and feedback from the second ANN (output from D is used in the objective function used to train G) to generate an artificial dataset (G generates artificial dataset pg).
and training the second ANN, using a plurality of artificial datasets generated by the ANN and a plurality of genuine datasets,
[pg. 4]
PNG
media_image8.png
681
1036
media_image8.png
Greyscale
NOTE: Teaches training the second ANN (D) using a plurality of artificial datasets (each iteration of the training process, a new artificial set of data is generated by the first ANN G, using G(zi), to train D), and a plurality of genuine datasets (each iteration of the training process, D is being trained using a different set of genuine data, xi)
to classify received datasets as artificial datasets or genuine datasets, and to provide feedback to the first ANN;
([pg. 2] We also define a second multilayer perceptron D(x;θd) that outputs a single scalar. D(x) represents the probability that x came from the data rather than pg. We train D to maximize the probability of assigning the correct label to both training examples and samples from G. We simultaneously train G to minimize log(1 − D(G(z))):)
NOTE: Teaches classifying received datasets as artificial datasets or genuine datasets (D assigns labels to data indicating if they are from the real or fake dataset, pg), and to provide feedback to the first ANN (G is trained using feedback from D, as the output of D is utilized in its objective function)
and outputting the trained ANN configured to convert the
([pg. 3] (d) After several steps of training, if G and D have enough capacity, they will reach a point at which both cannot improve because pg = pdata. The discriminator is unable to differentiate between the two distributions, i.e. D(x) = 1/2.)
NOTE: Teaches outputting the trained ANN configured to convert data to the generated dataset (the completion of training provides the trained ANN, G).
OBVIOUSNESS TO COMBINE GOODFELLOW, SHANKAR AND KILLORAN: Goodfellow is analogous art to the present disclosure as it pertains to generative adversarial neural networks.
The boson sampler of Shankar produces latent samples z, then a classical generator maps them to images. The generator of Goodfellow takes latent/noise vectors z and maps them to data, while a discriminator improves sample realism through adversarial feedback.
Additionally, Goodfellow states;
([pg. 7] Another advantage of adversarial networks is that they can represent very sharp, even degenerate distributions, while methods based on Markov chains require that the distribution be somewhat blurry in order for the chains to be able to mix between modes.)
NOTE: Discloses that a benefit of a GAN is that they are able to represent a larger variety of distributions.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to augment the ANN generator of claim 2 (as taught by Shankar in view of Killoran) using Goodfellows GAN framework to further expand the variety of distributions which the model is capable of representing.
Claim(s) 4-5, 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Tero Kerras et al. (hereinafter Kerras) (“A Style-Based Generator Architecture for Generative Adversarial Networks”, 03/29/2019)
Regarding claim 4, Shankar in view of Killoran teaches;
The method according to claim 2, wherein using the determined first set of latent vectors to train the ANN to convert the second set of one or more latent vectors to the generated dataset
(Using the same reasoning from the claim 2 rejection)
Shankar and Killoran fail to teach but Kerras teaches;
includes providing different latent vectors to different layers of the ANN.
[pg. 2]
PNG
media_image9.png
756
504
media_image9.png
Greyscale
([pg. 2] Comparing our approach to style transfer, we compute the spatially invariant style y from vector w instead of an example image.)
NOTE: Different latent vectors (w) are provided to different layers of the ANN (see figure 1).
OBVIOUSNESS TO COMBINE KERRAS WITH SHANKAR AND KILLORAN
Kerras is analogous art to Killoran and the present disclosure as it pertains to a generative ANN architecture, and Kerras is analogous art to Shankar as it pertains to generative models. Specifically, Kerras pertains to an architecture for a generative adversarial network which borrows from style transfer literature.
Additionally, Kerras states;
([Abstract] The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation proper ties, and also better disentangles the latent factors of variation.)
NOTE: This excerpt details that the proposed generative adversarial network architecture allows for better control, better structure in the latent space, and better data (image) quality.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to implement the generative adversarial model using the architecture proposed by Kerras to allow for better control, better structure in the latent space, and better data quality.
Regarding claim 5, Shankar in view of Killoran teaches;
The method according to claim 2,
(Using the same reasoning from the claim 2 rejection)
Shankar and Killoran fail to teach but Kerras teaches;
wherein the ANN comprises a convolutional neural network.
[pg. 2]
PNG
media_image9.png
756
504
media_image9.png
Greyscale
NOTE: The aforementioned ANN (the pictured style based generator) contains convolutional layers, and is therefore a convolutional neural network.
Regarding claim 8, Shankar in view of Killoran teaches;
The method according to claim 2,
(Using the same reasoning from the claim 2 rejection)
Shankar and Killoran fail to teach but Kerras teaches;
wherein determining, from the set of integer sequences, the first set of latent vectors comprises truncating
[pg. 8]
PNG
media_image10.png
302
324
media_image10.png
Greyscale
NOTE: Teaches determining (drawing) the latent vectors (w) from a truncated sample space.
OBVIOUSNESS:
In the disclosure of Shankar, the latent vectors are a drawn from a latent variable (sample space) Z, representing the integer sequences. Kerras discloses truncating their sample space, W, then determining latent vectors (w) from this sample space. By performing the process disclosed by Kerras in the context of Shankar, the integer sequences of the set of integer sequences within the latent variable (sample space) Z, would be truncated, and the resulting first set of latent vectors (z) would be determined from the truncated integer sequences of the set of integer sequences.
Kerras further states that;
[pg. 8]
PNG
media_image10.png
302
324
media_image10.png
Greyscale
NOTE: Discloses that drawing latent vectors from a truncated sampling space improves average quality of generated images.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to draw the latent vectors of claim 2 from a truncated sampling space to improve the quality of generated images.
Claim(s) 6 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Lars S. Madsen et al. (hereinafter Madsen) (“Quantum computational advantage with a programmable photonic processor”, 06/01/2022).
Regarding claim 6, Shankar in view of Killoran teach;
The method according to claim 2,
(Using the same reasoning from the claim 2 rejection)
Shankar and Killoran fail to teach but Madsen teaches; further comprising: selecting a plurality of parameter values to configure an interferometer of the boson sampler; wherein controlling the boson sampler comprises controlling the configured boson sampler.
([pg. 76] Fig. 1 | High-dimensional GBS from a fully programmable photonic processor. A periodic pulse train of single-mode squeezed states from a pulsed OPO enters sequence of three dynamically programmable loop-based interferometers. Each loop contains a VBS, including a programmable phase shifter, and an optical fibre delay line... Each run of the device involves the specification of 1,296 real parameters, corresponding to the sequence of settings for all VBS units.)
NOTE: Teaches selecting a plurality of parameter values to configure the interferometer of a boson sampler (GBS stands for Gaussian Boson Sampling), because each run specifies 1,296 real parameters for the programmable interferometer units, including programmable phase shifters and beam-splitting settings. The boson sampler is then controlled in that configured state to sample from the programmed distribution.
OBVIOUSNESS TO COMBINE MADSEN WITH SHANKAR AND KILLORAN:
Madsen is analogous art to Shankar, Killoran, and the present disclosure as it pertains to quantum computing. Specifically, Madsen pertains to quantum computational advantage using a programmable photonic processor.
Madsen additionally states;
([Abstract] Here we report quantum computational advantage using Borealis, a photonic processor offering dynamic programmability on all gates implemented. We carry out Gaussian boson sampling (GBS) on 216 squeezed modes entangled with three-dimensional connectivity, using a time-multiplexed and photon-number resolving architecture. On average, it would take more than 9,000 years for the best available algorithms and supercomputers to produce, using exact methods, a single sample from the programmed distribution, whereas Borealis requires only 36 μs. This runtime advantage is over 50 million times as extreme as that reported from earlier photonic machines.)
NOTE: This excerpt discloses that the programmable photonic processor of the disclosure of Madsen allows for a runtime advantage over earlier photonic machines.
Additionally, a photonic processor is the kind of hardware platform you would naturally use for variational boson sampling.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to implement the method of claim 2 (as taught by Shankar and Killoran) on the programmable photonic processor disclosed by Madsen to allow for competitive runtime of the system.
Claim(s) 11-12, 19-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Henggang Cui (Hereinafter Cui) (“Scalable deep learning on distributed GPUs with a GPU-specialized parameter server”, 2015).
Regarding claim 11,
Claim 11 is a system claim directly corresponding to claim 1 (taught by Shankar in view of Killoran), with the exception of the following limitation, which Shankar and Killoran fail to teach but Cui teaches;
A system comprising: … a set of one or more processors, the set of one or more processors configured to:
([Abstract] Large-scale deep learning requires huge computational resources to train a multi-layer neural network… This paper describes a new parameter server, called GeePS, that supports scalable deep learning across GPUs distributed among multiple machines)
NOTE: Teaches a set of processors (a set of GPUs indicates a plurality of processors, i.e. a set) for performing deep learning.
OBVIOUSNESS TO COMBINE CUI WITH SHANKAR AND KILLORAN
Cui is analogous art to the present disclosure as it pertains to deep learning methods. Specifically, Cui pertains to a scalable deep learning method on distributed GPUs with a specialized parameter server.
Additionally, Cui states;
([Abstract] Moreover, GeePS achieves the same training throughput with four GPU machines that a state-of-the-art CPU-only system achieves with 108 machines.)
NOTE: This excerpt discloses that the set of processors used by CUI greatly improved the training throughput of the deep learning.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to use the set of processors disclosed by Cui to implement the methods of claim 11 to improve the throughput of the deep learning system.
Regarding claim 12,
Claim 12 is a system claim directly corresponding to claim 2 (taught by Shankar in view of Killoran), with the exception of the following limitation, which is taught by Cui;
The system according to claim 11, wherein the set of one or more processors is configured to:
([Abstract] Large-scale deep learning requires huge computational resources to train a multi-layer neural network… This paper describes a new parameter server, called GeePS, that supports scalable deep learning across GPUs distributed among multiple machines)
NOTE: Teaches a set of processors (a set of GPUs indicates a plurality of processors, i.e. a set) for performing deep learning.
Regarding claim 19, Shankar in view of Killoran and Cui teaches;
The system according to claim 11,
(Using the same reasoning from claim 11)
Shankar fails to teach but Killoran teaches;
wherein the boson sampler is a Gaussian boson sampler.
([pg. 9] We will now overview how specific models of photonic quantum computing can be realized using CV quantum neural networks. First, Gaussian boson sampling (GBS) [106] isa model of photonic quantum computing where a multimode Gaussian state is prepared and subsequently measured in the photon-number basis, a procedure which is believed to be hard to simulate classically. GBS includes conventional Boson Sampling as a special case [107,108]. Any GBS configuration can be encoded in a CV quantum neural network by (i) turning off the non-Gaussian gates and (ii) measuring the outputs using photon detectors.)
NOTE: Teaches the boson sampler being a Gaussian boson sampler
Regarding claim 20, Shankar in view of Killoran and Cui teaches;
The system according to claim 11,
(Using the same reasoning from claim 11)
Shankar and Killoran fail to teach but Cui teaches;
wherein a processor of the set of processors comprises a graphics processing unit (GPU).
([Abstract] This paper describes a new parameter server, called GeePS, that supports scalable deep learning across GPUs distributed among multiple machines)
NOTE: Teaches a processor of the aforementioned set of processors comprises a graphics processing unit (GPU).
Claim(s) 13-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Cui (“Scalable deep learning on distributed GPUs with a GPU-specialized parameter server”, 2015), further in view of Madsen (“Quantum computational advantage with a programmable photonic processor”, 06/01/2022).
Regarding claim 13, Shankar in view of Killoran further in view of Cui teach;
The system according to claim 11,
(Using the same reasoning from the claim 11 rejection)
Shankar and Killoran fail to teach but Cui teaches;
the set of one or more processors is configured to:
(Using the same reasoning from the claim 11 rejection)
Shankar, Killoran, and Cui fail to teach but Madsen teaches;
wherein the boson sampler comprises a configurable interferometer, and wherein
([pg. 76] Fig. 1 | High-dimensional GBS from a fully programmable photonic processor. A periodic pulse train of single-mode squeezed states from a pulsed OPO enters sequence of three dynamically programmable loop-based interferometers. Each loop contains a VBS, including a programmable phase shifter, and an optical fibre delay line... Each run of the device involves the specification of 1,296 real parameters, corresponding to the sequence of settings for all VBS units.)
NOTE: Teaches selecting a plurality of parameter values to configure the interferometer of a boson sampler (GBS stands for Gaussian Boson Sampling), because each run specifies 1,296 real parameters for the programmable interferometer units, including programmable phase shifters and beam-splitting settings. The boson sampler is then controlled in that configured state to sample from the programmed distribution.
OBVIOUSNESS TO COMBINE MADSEN WITH SHANKAR, KILLORAN, CUI:
Madsen is analogous art to Shankar, Killoran, and the present disclosure as it pertains to quantum computing, and is analogous to Cui as it pertains to computer hardware. Specifically, Madsen pertains to quantum computational advantage using a programmable photonic processor.
Madsen additionally states;
([Abstract] Here we report quantum computational advantage using Borealis, a photonic processor offering dynamic programmability on all gates implemented. We carry out Gaussian boson sampling (GBS) on 216 squeezed modes entangled with three-dimensional connectivity, using a time-multiplexed and photon-number resolving architecture. On average, it would take more than 9,000 years for the best available algorithms and supercomputers to produce, using exact methods, a single sample from the programmed distribution, whereas Borealis requires only 36 μs. This runtime advantage is over 50 million times as extreme as that reported from earlier photonic machines.)
NOTE: This excerpt discloses that the programmable photonic processor of the disclosure of Madsen allows for a runtime advantage over earlier photonic machines.
Additionally, a photonic processor is the kind of hardware platform you would naturally use for variational boson sampling.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to implement the method of claim 11 (as taught by Shankar, Killoran, and Cui) using the programmable photonic processor disclosed by Madsen to allow for competitive runtime in the system.
Regarding claim 14, Shankar in view of Killoran further in view of Cui teach;
The system according to claim 11,
(Using the same reasoning from the claim 11 rejection)
Shankar, Killoran, and Cui fail to teach but Madsen teaches;
wherein the one or more photodetectors of the boson sampler are one or more photon number resolving (PNR) detectors.
[pg. 76]
PNG
media_image11.png
407
926
media_image11.png
Greyscale
(Fig. 1 | High-dimensional GBS from a fully programmable photonic processor. A periodic pulse train of single-mode squeezed states from a pulsed OPO enters a sequence of three dynamically programmable loop-based interferometers. Each loop contains a VBS, including a programmable phase shifter, and an optical fibre delay line. At the output of the interferometer, the Gaussian state is sent to a 1-to-16 binary switch tree (demux), which partially demultiplexes the output before readout by PNRs.)
NOTE: Teaches one or more photodetectors of a boson sampler being PNR detectors.
Using the same rational to combine Madsen as in claim 11, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to the PNR detectors as the one or more photodetectors of the system of claim 11 (as taught by Shankar, Killoran, and Cui).
Regarding claim 15, Shankar in view of Killoran, Cui and Madsen teach;
The system according to claim 14,
(Using the same reasoning from the claim 14 rejection)
Shankar, Killoran, and Cui fail to teach but Madsen teaches;
wherein each integer sequence of the one or more integer sequences is representative of a number of photons measured by a photodetector of the boson sampler.
([pg. 76] High-dimensional GBS from a fully programmable photonic processor. A periodic pulse train of single-mode squeezed states from a pulsed OPO enters a sequence of three dynamically programmable loop-based interferometers. Each loop contains a VBS, including a programmable phase shifter, and an optical fibre delay line. At the output of the interferometer, the Gaussian state is sent to a 1-to-16 binary switch tree (demux), which partially demultiplexes the output before readout by PNRs. The resulting detected sequence of 216 photon numbers, in approximately 36 μs, comprises one sample.)
NOTE: Teaches wherein each integer sequence of the one or more integer sequences is representative of a number of photons measured by a photodetector of the boson sampler. (each sample is an integer sequence representing a sequence of 216 photon number measured by the PNR of the boson sampler, and a sample is collected each loop of the iterative process, indicating one or more integer sequences).
OBVIOUSNESS:
The integer sequences of Madsen represent a number of photons measured by a photodetector of the boson sampler.
Additionally, the aforementioned integer sequences of Shankar also represent a number of measured photons:
[pg. 2]
PNG
media_image12.png
203
492
media_image12.png
Greyscale
From this, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, for the integer sequences disclosed by Shankar to have been collected using the PNR detectors of the boson sampler taught by Madsen.
Claim(s) 16, 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Cui (“Scalable deep learning on distributed GPUs with a GPU-specialized parameter server”, 2015), further in view of Nicolas Quesada et al. (hereinafter Quesada) (“Gaussian Boson Sampling using threshold detectors”, 12/18/2018).
Regarding claim 16, Shankar in view of Killoran, and Cui teach;
The system according to claim 11,
(Using the same reasoning as the claim 11 rejection)
Shankar, Killoran, and Cui fail to teach but Quesada teaches;
wherein the photodetectors are on/off detectors configured to indicate the presence and/or absence of photons.
NOTE: From [0075] of the applicant spec; “In other examples, the photodetectors may comprise threshold detectors, also known as on/off detectors.”
(Quesada [pg. 1] Although many of the models listed above have lessened the experimental difficulties of building a Boson Sampler, none of them has looked at what is perhaps the most experimentally accessible configuration: squeezed states undergoing linear operations sampled with threshold detectors. These binary outcome detectors measure whether there were 0 photons or 1 or more photons in the field being measured. As opposed to currently available PNRs based on superconducting technology, threshold detectors are inexpensive, commercially available, and can be operated at room temperature [20].)
NOTE: Teaches photodetectors of a boson sampler being on/off detectors (threshold detectors) configured to indicate the presence and/or absence of photons.
OBVIOUSNESS TO COMBINE QUESADA, SHANKAR, KILLORAN, AND CUI:
Quesada is analogous art to the present disclosure as it pertains to boson sampling using photodetectors. Specifically, they explore a boson sampling architecture which uses gaussian states sampled with threshold detectors.
Quesada additionally states;
([pg. 1] As opposed to currently available PNRs based on superconducting technology, threshold detectors are inexpensive, commercially available, and can be operated at room temperature [20].)
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to use the threshold detector described by Quesada as an alternative photodetector in the boson sampling architecture described in claim 11 (as taught by Shankar in view of Killoran and Cui) for a cheaper, more commercially available implementation which can be operated at room temperature.
Regarding claim 17, Shankar in view of Killoran, Cui, and Quesada teach;
The system according to claim 16,
(Using the same reasoning as the claim 16 rejection)
Shankar teaches;
wherein the one or more integer sequences comprises a set of binary strings,
[pg. 4]
PNG
media_image7.png
736
468
media_image7.png
Greyscale
NOTE: Teaches aforementioned one or more integer sequences comprising a binary vector (x) containing a set of binary strings (x1, x2, etc. represent binary strings 0, 1).
Shankar, Killoran and Cui fail to teach but Quesada teaches;
and wherein each binary integer of a binary string is representative of a presence or absence of photons in an output mode measured by a photodetector of the boson sampler.
(Although many of the models listed above have less ened the experimental difficulties of building a Boson Sampler, none of them has looked at what is perhaps the most experimentally accessible configuration: squeezed states undergoing linear operations sampled with threshold detectors. These binary outcome detectors measure whether there were 0 photons or 1 or more photons in the field being measured. As opposed to currently available PNRs based on superconducting technology, threshold detectors are inexpensive, commercially available, and can be operated at room temperature [20])
NOTE: Teaches each binary integer of a binary string (0 and 1 are considered binary integers of a binary string) is representative of a presence or absence of photons in an output mode (determine whether there were 0 photons [absence] or 1 or more photons [presence] in the field) measured by a photodetector of the boson sampler (measured by threshold detectors [which are photodetectors] of the boson sampler).
Claim(s) 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Shankar (“Neural Variational Boson Sampling”, 09/01/2022) in view of Killoran (“Continuous-variable quantum neural networks”, 10/31/2019) further in view of Cui (“Scalable deep learning on distributed GPUs with a GPU-specialized parameter server”, 2015), further in view of Yuxuan Li et al. (Hereinafter Li) (“Benchmarking 50-Photon Gaussian Boson Sampling on the Sunway TaihuLight”, June-2022).
Regarding claim 18, Shankar in view of Killoran and Cui teach;
The system according to claim 11,
(Using the same reasoning as the claim 11 rejection)
Shankar, Killoran, and Cui fail to teach but Li teaches;
wherein the boson sampler is a single-photon boson sampler.
NOTE: A single-photon boson sampler is being interpreted using the following excerpt from the applicant spec; “[0005] The boson sampler may comprise a single-photon boson sampler. In other words, the boson sampler may generate a photonic state comprising N single photons distributed across M modes of electromagnetic radiation (with N less than or equal to M), may provide that photonic state to an interferometer / optical network, and may measure the quantum superposition state output from the interferometer.”
(Li [pg. 2] In a typical boson sampling experiment, as shown in Fig. 1a, N indistinguishable single photons are sent to an M-port linear optical network, and the output scattered photons are detected by N single-photon detectors.)
PNG
media_image13.png
422
811
media_image13.png
Greyscale
NOTE: Teaches a single-photon boson sampler.
OBVIOUSNESS TO COMBINE LI WITH SHANKAR, KILLORAN, AND CUI: Li is analogous art to the present disclosure as it pertains to boson sampling. Specifically, Li pertains to the benchmarking of Gaussian Boson sampling with threshold detection based on a specific supercomputer.
Additionally, Li states;
([pg. 2] In a typical boson sampling experiment, as shown in Fig. 1a, N indistinguishable single photons are sent to an M-port linear optical network, and the output scattered photons are detected by N single-photon detectors.)
NOTE: This excerpt details that the aforementioned single-photon boson sampler is ‘typical’, i.e. a standard or conventional setup for boson sampling.
Additionally, Shankar already relies on integer valued boson sampler outputs.
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, for the single photon boson sampler implementation described by Li to be a predictable implementation choice for the boson sampling system of claim 11 (as taught by Shankar in view of Killoran and Cui).
Response to Arguments
In review of Applicant’s amendments, filed 06/12/2026, the abstract idea rejections from the previous office action made under 35 U.S.C. 101 have been withdrawn.
Applicant's arguments filed 06/12/2026 regarding the rejections under 35 U.S.C. 103 have been fully considered but they are not persuasive.
Starting on page 1, the applicant states that; the combination of Shankar and Killoran does not teach or suggest "controlling a boson sampler to produce one or more integer sequences, each of the one or more integer sequences determined from a measurement outcome of one or more photodetectors of the boson sampler," as recited in amended claim 1 because Shankar fails to teach or suggest "photodetectors" (see page 8) and "controlling a boson sampler to produce one or more integer sequences, each of the one or more integer sequences determined from a measurement outcome of...the boson sampler," as recited in amended claim 1 because Shankar does not describe using a physical boson sampler. Instead, Shankar describes simulating a boson sampler. More specifically, Shankar describes using a classical neural network model to generate a boson-sampling probability distribution. Due to this, the latent vectors of Shankar are not determined from measurement outcomes of an actual (physical) boson sampler. Instead, the latent vectors are determined from a probability distribution generated by a classical neural network model. Killoran does not remedy this deficiency of Shankar.
The applicant’s argument is not persuasive because Shankar teaches obtaining samples by measuring the output state of a boson-sampling circuit as an integer vector identifying the number of bosons in a corresponding output mode (as indicated in the previous and current office action). Shankar’s use of simulation to evaluate the disclosed system does not negate Shankar’s disclosure. Shankar is used for the express teaching of integer sequences determined from boson-sampler measurement outcomes, while Killoran is relied upon for expressly teaching that the measurements are performed using one or more photodetectors.
The applicant further states; the combination of Shankar and Killoran does not teach or suggest an ANN comprising "one or more hidden layers implementing nonlinear activation functions," as recited in amended claim 1. Killoran describes a "continuous-variable quantum neural network" implemented using variational quantum circuits composed of Gaussian and non-Gaussian quantum gates. As Killoran explains, the quantum neural network is a variational quantum circuit built in the continuous-variable (CV) architecture. Killoran's CV quantum neural network uses quantum gates (not classical hidden layers with nonlinear activation functions) to perform transformations. Accordingly, Killoran does not teach or suggest an ANN comprising "one or more hidden layers implementing nonlinear activation functions" that receives latent vectors determined from measurement outcomes of a boson sampler.
The applicant’s argument is not persuasive because Killoran expressly teaches the quantum neural network including one or more hidden layers implementing non-linear activation functions, as reflected by the 103 rejections for amended claim 1 in the current office action.
The applicant further states; Moreover, a person of ordinary skill in the art would not be motivated to modify the intentionally light-weight linear Gaussian decoder of Shankar using the continuous-variable quantum variational circuit architecture described by Killoran. Doing so would fundamentally alter the operating principles and computational assumptions underlying Shankar. Shankar expressly seeks to preserve simplicity and tractability by using a linear Gaussian decoder, and replacing this with a nonlinear multilayer architecture would undermine those design goals (see, e.g., Shankar, Section IV.B).
The applicant’s argument is not persuasive because it is presented in a conclusory manner. The applicant states that the modification would “fundamentally alter” unspecified “computational assumptions,” which is not, by itself, an explanation of inoperability or a changed operating principle. Even assuming arguendo that the combination allegedly undermines design goals including simplicity and tractability, loss of simplicity and tractability does not defeat motivation. Shankar itself expressly identifies better sample quality as a reason to use a more complex decoder ([Shankar, pg. 5] One can use more complex decoder models for better sample quality).
For the above reasons, the rejections under 35 U.S.C. 103 regarding the amended independent claims, as well as the claims that depend therefrom, stand.
CONCLUSION
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Matthew Alan Cady whose telephone number is (571) 272-7229. The examiner can normally be reached Monday - Friday, 7:30 am - 5:00 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached on (571)272-4128. 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.
/MATTHEW ALAN CADY/ Examiner, Art Unit 2145
/CESAR B PAULA/ Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights