Granted Patent B2

US 12,423,374 · App. 16/888,419 · Granted Sep 23, 2025

Inventors: Michael Paul Friedlander (Vancouver, CA); Pooya Ronagh (Vancouver, CA); Behrooz Sepehry (Vancouver, CA)

Assignee: 1QB INFORMATION TECHNOLOGIES INC.

G06F17/11G06F17/18G06F18/211G06F18/213G06N3/02G06N7/08G06N20/10

Abstract

Claims (48)

1. A method for training a machine learning model based at least in part on solving an optimization problem using a digital computer and a sampling device, said method comprising:

(a) receiving, by said digital computer, data corresponding to said optimization problem, said data comprising one or more objective functions, wherein an objective function of said one or more objective functions is configured to accept at least an outer optimization argument of said objective function and an inner optimization argument of said objective function, wherein said outer optimization argument is representative of model trainable parameters of said machine learning model; and

(b) training, by said digital computer, said machine learning model at least in part by:

(i) providing values of said inner optimization argument of said objective function to said sampling device, wherein said sampling device comprises at least one non-classical computer, and wherein said sampling device is configured to generate one or more samples based at least in part on said inner optimization argument;

(ii) receiving said one or more samples from said sampling device;

(iii) using said one or more samples to perform an inner optimization over said inner optimization argument of said objective function to generate an inner optimization result;

(iv) determining one or more gradients or sub-gradients of said objective function based at least in part on said inner optimization result;

(v) performing one or more classical optimization protocols based at least in part on said one or more gradients or sub-gradients to thereby obtain an outer optimization over said outer optimization argument;

(vi) outputting an outer optimization result for said machine learning model; and

(vii) updating said model trainable parameters based at least in part on said outer optimization result to train said machine learning model.

2. The method of claim 1 , further comprising repeating (b) in an iterative manner until a stopping criterion is met.

3. The method of claim 1 , wherein said one or more objective functions comprise a loss function.

4. The method of claim 1 , wherein said objective function comprises one or more composite functions of said outer optimization argument and said inner optimization argument.

5. The method of claim 1 , wherein (b)(iv) is performed by one or more applications of a chain rule.

6. The method of claim 5 , wherein said one or more applications comprise one or more iterative applications of said chain rule.

7. The method of claim 6 , wherein said one or more applications are performed using auto-differentiation.

8. The method of claim 4 , wherein one or more argument functions of said one or more composite functions comprises one or more differentiable feature extractors.

9. The method of claim 8 , wherein said one or more differential feature extractors comprise one or more deep neural networks.

10. The method of claim 1 , wherein said data further comprises a set of permissible values for said inner optimization argument.

11. The method of claim 10 , wherein (b) comprises using said sampling device to generate said one or more samples from said set of permissible values.

12. The method of claim 10 , further comprising generating said one or more samples based at least in part on a probability distribution.

13. The method of claim 12 , wherein said probability distribution is determined at least in part by said objective function and a set of scaling parameters.

14. The method of claim 1 , wherein (b)(v) comprises applying one or more members selected from the group consisting of a stochastic gradient descent (SGD) method, a stochastic average gradient (SAG or SAGA) method, a stochastic variance-reduced gradient descent (SVRG) method, a stochastic dual coordinate ascent (SDCA) method, an adaptive moment estimation (Adam) method, a reduced mean square (RMS) method, a root mean square propagation (RMSProp) method, and an adaptive gradient algorithm (AdaGrad).

15. The method of claim 1 , wherein said at least one non-classical computer comprises at least one quantum computer.

16. The method of claim 1 , wherein said at least one non-classical computer comprises at least one quantum annealer.

17. The method of claim 1 , wherein said digital computer is remotely-located with respect to said sampling device.

18. The method of claim 1 , wherein said sampling device comprises one or more optical parametric oscillators.

19. The method of claim 1 , wherein said optimization problem comprises a non-convex problem.

20. The method of claim 1 , wherein said sampling device comprises reconfigurable digital hardware configured to perform a Markov chain Monte Carlo procedure.

21. The method of claim 20 , wherein said Markov chain Monte Carlo procedure comprises one or more members selected from the group consisting of a simulated annealing procedure, a simulated quantum annealing procedure, and a Gibbs sampling procedure.

22. The method of claim 1 , wherein said optimization problem is associated with training a structured support vector machine (SVM).

23. The method of claim 1 , wherein said data corresponding to said optimization problem is associated with one or more members selected from the group consisting of an image segmentation problem, a compressed sensing problem, a basis pursuit problem from a compressed sensing problem, a dual of a basis pursuit problem from a compressed sensing problem, a semi-supervised learning problem, a noun phrase co-reference resolution problem, an active learning problem, an image tagging problem, and a recommender system.

24. The method of claim 1 , wherein said sampling device further comprises at least one digital computer, field-programmable gate array, graphics processing unit, or application-specific integrated circuit.

25. The method of claim 1 , further comprising using said outer optimization result of said outer optimization for performing one or more operations in an iterative optimization process in machine learning.

26. A system for training a machine learning model based at least in part on solving an optimization problem, comprising a digital computer and a sampling device, wherein said digital computer is configured to at least:

(a) receive data corresponding to said optimization problem, said data comprising one or more objective functions, wherein an objective function of said one or more objective functions is configured to accept at least an inner optimization argument and an outer optimization argument, wherein said outer optimization argument is representative of model trainable parameters of said machine learning model; and

(b) train said machine learning model at least in part by:

(i) providing values of said inner optimization argument of said objective function to said sampling device, wherein said sampling device comprises at least one non-classical computer, and wherein said sampling device is configured to generate one or more samples based at least in part on said inner optimization argument,

(ii) obtaining, via an inner optimization, an inner optimization result over said inner optimization argument of said objective function from said sampling device,

(iii) determining one or more gradients or sub-gradients of said objective function based at least in part on said inner optimization,

(iv) performing one or more classical optimization protocols based at least in part on said one or more gradients or sub-gradients to thereby obtain an outer optimization result over said outer optimization argument,

(v) output said outer optimization result for said machine learning model, and

(vi) updating said model trainable parameters based at least in part on said outer optimization result to train said machine learning model.

27. The method of claim 1 , wherein said one or more samples comprise samples of discrete vectors distributed according to a Boltzmann distribution.

28. The method of claim 1 , wherein performing said inner optimization in (b) (iii) uses a smoothed maximum function.

29. The system of claim 26 , wherein said digital computer is further configured to use said outer optimization result of said outer optimization for performing one or more operations in an iterative optimization process in machine learning.

30. The system of claim 29 , wherein said sampling device further comprises at least one digital computer, field-programmable gate array, graphics processing unit, or application-specific integrated circuit.

31. The system of claim 29 , wherein said non-classical computer comprises at least one quantum computer or at least one quantum annealer.

Assignments (1)

Continuity (4)

References Cited (400)