Patent US10229355

Quantum processor and its use for implementing a neural network
[+] Show Details

US10229355

1. A quantum processor comprising: (1) (7)

a first group of substantially parallel superconducting quantum circuits;

a second group of substantially parallel superconducting quantum circuits, wherein the second group is substantially perpendicular to the first group; and

a third group of substantially parallel superconducting quantum circuits, wherein the third group is substantially perpendicular to the second group;

wherein the quantum processor has a non-overlapping layout, such that:

each superconducting quantum circuits of tire first group is sized and shaped so that it crosses only the second group,

each superconducting quantum circuit of the second group is sized and shaped so that it crosses only the first group and the third group, and

each superconducting quantum circuits of the third group is sized and shaped so that it crosses only the second group and a successive group of parallel superconducting quantum circuits which is perpendicular to the third group.

2. The quantum processor as claimed in claim 1, wherein the quantum processor is used for implementing a neural network comprising a plurality of neurons and a plurality of synapses; wherein each neuron of the plurality of neurons is associated to a qubit and each synapse of the plurality of synapses is associated to a coupler of the quantum processor. (4) (0)

3. A method for training the neural network implemented in the quantum processor claimed in claim 2, the method comprising: (4) (7)

providing initialization data for initializing the plurality of couplers and the superconducting quantum circuits of the quantum processor;

until a criterion is met:

performing a quantum sampling of the quantum processor to provide first empirical means;

obtaining at least one training data instance for training the neural network;

performing a quantum sampling of the quantum processor; wherein no bias is assigned to the superconducting quantum circuits of the first group: wherein couplings of the first group and the second group are switched off; further wherein the biases of the second group are altered using the biases on a first group of neurons associated with the first group of superconducting quantum circuits, the weights of the switched off couplings, and the at least one training data instance, to determine second empirical means;

updating corresponding weights and biases of the couplers and the superconducting quantum circuits of the quantum processor using the first and second empirical means;

and providing final weights and biases of the couplers and the superconducting quantum circuits of the quantum processor indicative of data representative a trained neural network.

4. The method as claimed in claim 3, wherein the initialization data comprise a plurality of biases, each for a superconducting quantum circuits of the quantum processor; (1) (1)

a plurality of weights, each weight for a coupler of the plurality of couplers and a learning rate schedule.

5. The method as claimed in claim 4, wherein the providing of the initialization data is performed using an analog computer comprising the quantum processor and a digital computer operatively connected to the analog computer. (2) (0)

8. The method as claimed in claim 3, wherein the at least one training data instance is obtained from a previously generated data set. (1) (0)

10. The method as claimed in claim 3, wherein the at least one training data instance is obtained from a real-time source. (1) (0)

12. The method as claimed in claim 3, wherein the criterion comprises a stopping condition; wherein the stopping condition comprises determining if there is no further training data instance available. (0)

13. The method as claimed in claim 2, wherein the neural network is a Deep Boltzmann Machine. (0)

14. A digital computer comprising: (0) (11)

a central processing unit;

a display device;

a communication port for operatively connecting the digital computer to an analog computer comprising a quantum processor used for implementing a neural network as claimed in claim 2;

a memory unit comprising an application for training the neural network, the application comprising:

instructions for providing initialization data for initializing die plurality of couplers and the plurality of biases of the superconducting quantum circuits of the quantum processor;

instructions for, until a criterion is met:

performing a quantum, sampling of the quantum processor to provide first empirical means;

obtaining at least one training data instance for training the neural network;

performing a quantum sampling of the quantum processor; wherein no bias is assigned to the superconducting quantum circuits of the first group; wherein the couplings of the first group and the second group are switched off; farther wherein the biases of the second group of superconducting quantum circuits are altered using the biases on a first group of neurons associated with the first group of superconducting quantum circuits, the weights of the switched off couplings, and the at least one training data instance, to determine second empirical means; and

updating corresponding weights and biases of the couplers and the superconducting quantum circuits of the quantum processor using the first and the second empirical means; and

instructions for providing final weights and biases of the couplers and the superconductive quantum circuits of the quantum processor as data representative of a trained neural network.

15. A non-transitory computer readable storage medium for storing computer-executable instructions which, when executed, cause a digital computer to perform a method for training the neural network implemented in the quantum processor claimed in claim 2, the method comprising: (0) (7)

providing initialization data for initializing the plurality of couplers and the superconducting quantum circuits of the quantum processor;

until a criterion is met:

performing a quantum sampling of the quantum processor to provide first empirical means;

obtaining at least one training data instance for training the neural network;

performing a quantum sampling of the quantum processor; wherein no bias is assigned to the superconducting quantum circuits of the first group; wherein the couplings of the first group and the second group are switched off; further wherein the biases of the second group of superconducting quantum circuits are altered using the biases on a first group of neurons associated with the first group of superconducting quantum circuits, the weights of the switched off couplings, and the at least one training data instance, to determine second empirical means; and

updating corresponding weights and biases of the couplers and the superconducting quantum circuits of tire quantum processor using the first and second empirical means; and

providing final weights and biases of the couplers and the qubits of the quantum processor indicative of data representative a trained neural network.

View Abstract and Specification Size

PDF with Images and Document Face >

Full Text Publication >

Patent US10229355

Quantum processor and its use for implementing a neural network

Patent Matrix® Search