Granted Patent B1

US 12,406,188 · App. 17/193,812 · Granted Sep 2, 2025

Inventors: Santiago Gonzalez (Denver, CO); Jason Zhi Liang (Fremont, CA); Risto Miikkulainen (Stanford, CA)

Assignee: Cognizant Technology Solutions U.S. Corportion

G06N3/086G06F16/215G06N3/04

Abstract

Claims (30)

1. A process for evolving a data augmentation policy for application to sample data from a dataset, wherein the sample data is used to train a deep neural network (DNN) to perform a predetermined task, the process comprising:

evolving an initial population of candidate data augmentation policy models, wherein each initial candidate data augmentation policy model includes multiple nodes and multiple edges and further wherein each node represents a single distinct data augmentation operation, and each edge indicates a weight between two nodes, the weight representing a probability related to action by a second node on input data from a first node;

evaluating each initial candidate data augmentation policy model by:

i. applying each initial candidate data augmentation policy to the sample data to produce an augmented dataset;

ii. at least partially training the deep neural network (DNN) using the augmented sample dataset;

iii. determining a fitness for each initial candidate data augmentation policy model, wherein the candidate data augmentation policy model's fitness is accuracy of the at least partially trained deep neural network (DNN) on a held-out validation dataset from the dataset;

selecting one of (a) a final evaluated data augmentation policy or (b) one or more evaluated initial candidate data augmentation policy models for reproduction on the basis of determined fitness;

upon selection of (b), reproducing child candidate data augmentation policy models from the selected one or more evaluated initial candidate data augmentation policy models; and repeating evaluating and selecting for the child candidate data augmentation policy models until resulting selection is (a), wherein the initial population is separated into subpopulations of candidate data augmentation policy models in accordance with similarity between candidate topologies and each subpopulation is separately subjected to the evaluating, the selecting and the repeating;

applying the final evaluated data augmentation policy on sample data;

training the deep neural network using the sample data on which the final evaluated data augmentation policy has been applied.

2. The process according to claim 1 , wherein the single distinct data augmentation operations are selected from the group consisting of: pixel-wise arithmetic; downsampling; blurring; cropping; affine transforms; noise injection; contrast adjustment; and tonemapping.

3. The process according to claim 1 , wherein the probability related to action by a second node on input data from a first node indicates a probability that a second node operation is applied after a first node operation in a case where there is a single first node converges on a second node.

4. The process according to claim 1 , wherein the probability related to action by a second node on input data from a first node indicates a probability that an input from the first node will be used by the second node in a case where there are multiple first nodes converging on a single second node.

5. The process according to claim 1 , wherein the dataset includes data from one of the following domains imaging, text, time series, speech.

6. The process according to claim 1 , wherein the reproducing child candidate data augmentation policy models includes at least one of recombining and mutating the selected one or more evaluated initial candidate data augmentation policy models.

7. A non-transitory computer-readable medium storing instructions that, when executed by a computer, perform a process for evolving a data augmentation policy for application to sample data from a dataset, wherein the sample data is used to train a deep neural network (DNN) to perform a predetermined task, comprising:

evolving an initial population of candidate data augmentation policy models, wherein each initial candidate data augmentation policy model includes multiple nodes and multiple edges and further wherein each node represents a single distinct data augmentation operation, and each edge indicates a weight between two nodes, the weight representing a probability related to action by a second node on input data from a first node;

evaluating each initial candidate data augmentation policy model by:

i. applying each initial candidate data augmentation policy to the sample data to produce an augmented dataset;

ii. at least partially training the deep neural network (DNN) using the augmented sample dataset;

iii. determining a fitness for each initial candidate data augmentation policy model, wherein the candidate data augmentation policy model's fitness is accuracy of the at least partially trained deep neural network (DNN) on a held-out validation dataset from the dataset;

selecting one of (a) a final evaluated data augmentation policy or (b) one or more evaluated initial candidate data augmentation policy models for reproduction on the basis of determined fitness;

upon selection of (b), reproducing child candidate data augmentation policy models from the selected one or more evaluated initial candidate data augmentation policy models; and repeating evaluating and selecting for the child candidate data augmentation policy models until resulting selection is (a), wherein the initial population is separated into subpopulations of candidate data augmentation policy models in accordance with similarity between candidate topologies and each subpopulation is separately subjected to the evaluating, the selecting and the repeating;

applying the final evaluated data augmentation policy on sample data;

training the deep neural network using the sample data on which the final evaluated data augmentation policy has been applied.

8. The non-transitory computer readable medium according to claim 7 , wherein the single distinct data augmentation operations are selected from the group consisting of: pixel-wise arithmetic; downsampling; blurring; cropping; affine transforms; noise injection; contrast adjustment; and tonemapping.

9. The non-transitory computer readable medium according to claim 7 , wherein the probability related to action by a second node on input data from a first node indicates a probability that a second node operation is applied after a first node operation in a case where there is a single first node converges on a second node.

10. The non-transitory computer readable medium according to claim 7 , wherein the probability related to action by a second node on input data from a first node indicates a probability that an input from the first node will be used by the second node in a case where there are multiple first nodes converging on a single second node.

11. The non-transitory computer readable medium according to claim 6 , wherein the dataset includes data from one of the following domains imaging, text, time series, speech.

12. The non-transitory computer readable medium according to claim 7 , wherein the reproducing child candidate data augmentation policy models includes at least one of recombining and mutating the selected one or more evaluated initial candidate data augmentation policy models.

Assignments (1)

Continuity (1)

References Cited (178)