Executing a genetic algorithm on a low-power controller

Patent metadata

Field	Value
Publication number	US-12387105-B2
Application number	US-202318520192-A
Country	US
Kind code	B2
Filing date	Nov 27, 2023
Priority date	Oct 16, 2019
Publication date	Aug 12, 2025
Grant date	Aug 12, 2025

Abstract

A method includes selecting a subset of models from a plurality of models. The plurality of models is generated based on a genetic algorithm and corresponds to a first epoch of the genetic algorithm. Each of the plurality of models includes data representative of a neural network. The method includes performing at least one genetic operation of the genetic algorithm with respect to at least one model of the subset to generate a trainable model. The method includes determining a rate of improvement associated with prior backpropagation iterations. The method includes selecting, based on the rate of improvement, one of the trainable model or a prior trainable model as a selected trainable model. The method includes generating the trained model including training the selected trainable model. The method includes adding the trained model as input to a second epoch of the genetic algorithm that is subsequent to the first epoch.

First claim

What is claimed is:

1. A computer system comprising:
a memory configured to store a training data set and a plurality of data structures, each of the plurality of data structures including data representative of a neural network;
a processor configured to execute a recursive search, wherein executing the recursive search comprises, during a first iteration of the recursive search:
determining a fitness value for each of the plurality of data structures based on at least a subset of the training data set;
selecting a subset of data structures from the plurality of data structures based on the fitness values of the subset of data structures;
performing at least one of a crossover operation or a mutation operation with respect to at least one data structure of the subset to generate a trainable data structure; and
providing the trainable data structure to an optimization trainer; and

the optimization trainer configured to, during a particular backpropagation iteration:
determine a rate of improvement associated with a plurality of prior backpropagation iterations, the prior backpropagation iterations performed prior to the particular backpropagation iteration;
select, based on the rate of improvement, one of the trainable data structure or a prior trainable data structure as a selected trainable data structure, the prior trainable data structure previously trained during a first backpropagation iteration of the prior backpropagation iterations;
determine a learning rate based on a difference between an error measure and a non-zero error threshold, wherein a weight of a connection is updated based at least in part on the learning rate;
generate a trained data structure, including training the selected trainable data structure based on a portion of the training data set; and
provide the trained data structure as input to a second iteration of the recursive search that is subsequent to the first iteration of the recursive search.

2. The computer system of claim 1 , wherein the optimization trainer and the recursive search are executed on a single device.

3. The computer system of claim 2 , wherein the single device is integrated into a sensor of a plurality of sensors, and wherein the training data set is generated by the plurality of sensors.

4. The computer system of claim 1 , wherein the optimization trainer is executed on a different device, graphics processing unit (GPU), processor, core, thread, or any combination thereof, than the recursive search.

5. The computer system of claim 1 , wherein the optimization trainer is further configured to train the selected trainable data structure including:
determining a node error measure of a node of the selected trainable data structure;
determining an error measure based at least in part on the node error measure; and
selectively updating a weight of a connection of the node based on determining that the error measure is greater than a non-zero error threshold.

6. The computer system of claim 1 , wherein the first iteration of the recursive search is performed with a first set of weights, wherein the second iteration of the recursive search is performed with a second set of weights, and wherein the optimization trainer is further configured to generate a third set of weights for a third iteration of the recursive search prior to execution of the first iteration of the recursive search.

7. The computer system of claim 5 , wherein the optimization trainer is configured to determine the error measure based on node error measures of a plurality of nodes of the trainable data structure, and wherein the plurality of nodes includes the node.

8. The computer system of claim 5 , wherein the node is included in a layer of the trainable data structure, and wherein the error measure is based on node error measures of nodes included in the layer.

9. The computer system of claim 1 , wherein the processor is further configured to:
receive an input data point; and
based on determining that a difference between the input data point and a reference point satisfies a training threshold, add the input data point to the training data set.

10. The computer system of claim 9 , wherein the processor is further configured to determine the reference point based on a plurality of data points of the training data set, and wherein each of the input data point and the plurality of data points are associated with the same classifier result.

11. The computer system of claim 9 , wherein the difference between the input data point and the reference point corresponds to a distance between the input data point and the reference point in a vector space.

12. The computer system of claim 1 , wherein executing the recursive search further comprises, during a first iteration of the recursive search:
grouping the data structures of the plurality of data structures into species based on genetic distance between the data structures;
determining a species fitness value of each of the species; and
selectively removing one or more species from the crossover operation or the mutation operation responsive to determining that the one or more species satisfy a stagnation criterion.

13. A method comprising:
based on a first set of weights, executing, at a first device, one or more first iterations of an optimization process with respect to a model that includes data representative of a neural network;
determining, at the first device, a rate of improvement associated with the one or more first iterations of the optimization process;
responsive to the rate of improvement failing to satisfy a threshold, executing, at the first device, the optimization process on the model using a second set of weights that differs from the first set of weights;
executing, at the first device, one or more third iterations of the optimization process based on a third set of weights prior to executing the one or more first iterations; and
providing a processed neural network to a second device distinct from the first device, the processed neural network based on the second set of weights, wherein the second device is configured to perform functions utilizing the processed neural network without performing the optimization process on the processed neural network.

14. The method of claim 13 , wherein the optimization process includes determining a learning rate based on a difference between an error measure and a non-zero error threshold, and wherein a weight of a connection is updated based at least in part on the learning rate.

15. The method of claim 13 , wherein second set of weights are the same as the third set of weights.

16. The method of claim 13 , wherein the rate of improvement corresponds to a difference between a first result from executing the one or more first iterations and a second result from executing the one or more third iterations.

17. The method of claim 16 , wherein the difference corresponds to a particular difference in a number of errors in the first result and a third result from executing the one or more third iterations.

18. The method of claim 13 , wherein the optimization process comprises updating connection weights of the model but not a topology or activation functions of the model.

19. A non-transitory computer-readable storage device storing instructions that, when executed, cause a computer to perform operations comprising:
based on a first set of weights, executing one or more first iterations of an optimization process with respect to a model that includes data representative of a neural network, where

Assignees

• Sparkcognition Inc
• Avathon Inc

Inventors

• Husain Syed Mohammad Amir

Classifications

• G06N3/126

  Evolutionary algorithms, e.g. genetic algorithms or genetic programming · CPC title

• G06N3/063

  using electronic means · CPC title

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/082 Primary

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title