Evolution of deep neural network structures

Patent metadata

Field	Value
Publication number	US-11250327-B2
Application number	US-201715794905-A
Country	US
Kind code	B2
Filing date	Oct 26, 2017
Priority date	Oct 26, 2016
Publication date	Feb 15, 2022
Grant date	Feb 15, 2022

Abstract

The technology disclosed relates to evolving deep neural network structures. A deep neural network structure includes a plurality of modules with submodules and interconnections among the modules and the submodules. In particular, the technology disclosed relates to storing candidate genomes that identify respective values for a plurality of hyperparameters of a candidate genome. The hyperparameters include global topology hyperparameters, global operational hyperparameters, local topology hyperparameters, and local operational hyperparameters. It further includes evolving the hyperparameters by training, evaluating, and procreating the candidate genomes and corresponding modules and submodules.

First claim

What is claimed is:

1. A computer-implemented system for evolving a deep neural network structure, the deep neural network structure including a plurality of modules and interconnections among the modules, comprising:
a memory storing a candidate genome database having a pool of candidate genomes, each of the candidate genomes identifying respective values for a plurality of hyperparameters of the candidate genome, the hyperparameters including global topology hyperparameters identifying a plurality of modules in the genome and interconnections among the modules in the genome, at least one of the modules in each candidate genome including a neural network, the neural network including a plurality of layers, and the hyperparameters further including local topology hyperparameters identifying a plurality of submodules of the neural network and interconnections among the submodules, each candidate genome having associated therewith storage for an indication of a respective fitness value;
a candidate pool processor which:
trains the modules identified by the genome, including modifying the submodules of the neural network and their interconnections in dependence upon a predetermined back-propagation algorithm;
evaluates genomes from the candidate pool on validation data, including updating the fitness value associated with each of the genomes being evaluated;
selects genomes from the candidate pool for discarding in dependence upon their updated fitness values;
forms new genomes in dependence upon a respective set of at least one parent genome from the candidate pool; and
deploys a selected one of the genomes from the candidate pool.

2. The system of claim 1 , wherein the candidate pool processor further initializes the memory with an initial candidate genome pool.

3. The system of claim 2 , wherein global topology hyperparameters of each of the candidate genomes in the initial candidate genome pool identify a plurality of minimal structure modules in each candidate genome.

4. The system of claim 3 , wherein at least one of the minimal structure modules is a neural network with zero hidden submodules.

5. The system of claim 3 , wherein each of the candidate genomes in the initial candidate genome pool identifies uniform respective values for the global topology hyperparameters in each candidate genome.

6. The system of claim 3 , wherein each of the candidate genomes in the initial candidate genome pool identifies different respective values for the global topology hyperparameters in each candidate genome.

7. The system of claim 3 , wherein each of the candidate genomes in the initial candidate genome pool identifies different respective values for at least one of the local topology hyperparameters in each candidate genome.

8. The system of claim 3 , wherein each of the candidate genomes in the initial candidate genome pool identifies different respective values for at least one of local operational hyperparameters in each candidate genome.

9. The system of claim 3 , wherein each of the candidate genomes in the initial candidate genome pool identifies different respective values for at least one of global operational hyperparameters in each candidate genome.

10. The system of claim 3 , wherein the forming new genomes incrementally complexifies the minimal structure modules in each candidate genome.

11. The system of claim 10 , wherein the incremental complexification comprises adding new submodules in the minimal structure modules using mutation.

12. The system of claim 3 , wherein new genomes are formed in dependence upon a respective set of at least one parent genome with at least one minimal structure module, and wherein certain new genomes identify global topology hyperparameter values identifying new complex submodules formed in dependence upon the minimal structure module using crossover.

13. The system of claim 3 , wherein new genomes are formed in dependence upon a respective set of at least one parent genome with at least one minimal structure module, and, wherein at least one of the new genomes identifies values for global topology hyperparameters identifying new complex submodules formed in dependence upon the minimal structure module using crossover.

14. The system of claim 1 , wherein the modules identified by each of more than one of the candidate genomes include more than one neural network.

15. The system of claim 1 , wherein the modules identified by one of the candidate genomes include a convolutional neural network.

16. The system of claim 1 , wherein the modules identified by one of the candidate genomes include a convolution module.

17. The system of claim 1 , wherein the modules identified by one of the candidate genomes include a fully-connected neural network.

18. The system of claim 1 , wherein the global topology hyperparameters identify a type for each of the modules in the genome.

19. The system of claim 1 , wherein the global topology hyperparameters identify a sequence of processing data through each of the modules in the genome.

20. The system of claim 1 , wherein the global topology hyperparameters identify a branching and rejoining of modules in the genome.

21. The system of claim 20 , wherein the global topology hyperparameters identify an interconnection in one branch that skips over at least one module in another branch.

22. The system of claim 1 , wherein the local topology hyperparameters identifying submodules of the neural network include a number of neuron layers in the neural network.

23. The system of claim 1 , wherein the hyperparameters further include global operational hyperparameters that apply to entire genomes.

24. The system of claim 1 , wherein the hyperparameters further include local operational hyperparameters that are specific to respective modules in the genome.

25. The system of claim 1 , wherein the candidate pool processor further groups the evaluated genomes from the candidate pool into species by similarity, and wherein genomes are discarded in dependence upon their updated fitness values comprises comparing the updated fitness values of genomes only to other genomes in the same species.

26. A method of evolving a deep neural network structure, the deep neural network structure including a plurality of modules and interconnections among the modules, including:
storing a candidate genome database having a pool of candidate genomes, each of the candidate genomes identifying respective values for a plurality of hyperparameters of the candidate genome, the hyperparameters including global topology hyperparameters identifying a plurality of modules in the genome and interconnections among the modules in the genome, at least one of the modules in each candidate genome including a neural network, the neural network including a plurality of layers, and the hyperparameters further including local topology hyperparameters identifying a plurality of submodules of the neural network and interconnections among the submodules, each candidate genome having associated therewith storage for an indication of a respective fitness value;
training the modules identified by the genome, including modifying the submodules of the neural network and their interconnections in dependence upon a predetermined back-propagation algorithm;
evaluating genomes from the candidate pool on validation data, including updating the fitness value associated with each of the genomes be

Assignees

• Cognizant Tech Solutions U S Corporation

Inventors

• Liang Jason Zhi
• Miikkulainen Risto

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title