System and method for optimization of deep learning architecture

Patent metadata

Field	Value
Publication number	US-11017269-B2
Application number	US-201716334091-A
Country	US
Kind code	B2
Filing date	Jun 21, 2017
Priority date	Sep 30, 2016
Publication date	May 25, 2021
Grant date	May 25, 2021

Abstract

A method for determining optimized deep learning architecture includes receiving a plurality of training images and a plurality of real time images corresponding to a subject. The method further includes receiving, by a medical practitioner, a plurality of learning parameters comprising a plurality of filter classes and a plurality of architecture parameters. The method also includes determining a deep learning model based on the plurality of learning parameters and the plurality of training images, wherein the deep learning model comprises a plurality of reusable filters. The method further includes determining a health condition of the subject based on the plurality of real time images and the deep learning model. The method also includes providing the health condition of the subject to the medical practitioner.

First claim

The invention claimed is:

1. A method for determining optimized deep learning architecture, comprising:
receiving a plurality of training images and a plurality of real time images corresponding to a subject;
receiving, by a medical practitioner, a plurality of learning parameters comprising a plurality of filter classes and a plurality of architecture parameters;
determining a deep learning model based on the plurality of learning parameters and the plurality of training images, wherein the deep learning model comprises a plurality of filters, wherein the determining the deep learning model comprises reducing a number of the plurality of learning parameters of the deep learning model based at least on one or more constraints that determine a coefficient set that is applied to the plurality of filters, and wherein the coefficient set is determined based on at least a sparsity constraint and a convexity constraint applied to rows or columns of a parameter matrix;
determining a health condition of the subject based on the plurality of real time images and the deep learning model; and
providing the health condition of the subject to the medical practitioner.

2. The method of claim 1 , wherein determining the deep learning model comprises determining a convolution neural network, and wherein the plurality of filters comprise at least one reusable filter.

3. The method of claim 2 , wherein determining the health condition comprises:
receiving a plurality of input feature maps, wherein each input feature map of the plurality of input feature map is a representation of the plurality of real time images; and
generating a plurality of output feature maps based on the convolution neural network, wherein each of the plurality of output feature map is a linear combination of filtered version of the plurality of input feature maps, wherein each of the plurality of output feature map is a representation of the input feature maps.

4. The method of claim 3 , wherein determining the deep learning model comprises determining a plurality of seed filters, wherein dimensionality of the seed filters is based on a size of the plurality of training images.

5. The method of claim 4 , wherein determining the deep learning model comprises selecting a plurality of optimal filters among the plurality of seed filters for generating an output feature map of the plurality of output feature maps.

6. The method of claim 4 , wherein determining the deep learning model comprises determining the coefficient set corresponding to each output feature map of the plurality of output feature maps based on the plurality of seed filters.

7. The method of claim 4 , wherein determining the deep learning model comprises determining a dictionary of optimal filters for generating the plurality of output feature maps based on the plurality of seed filters.

8. The method of claim 7 , wherein determining the deep learning model comprises determining the coefficient set corresponding to each output feature map of the plurality of output feature maps based on the dictionary of optimal filters.

9. The method of claim 4 , wherein determining the deep learning model comprises determining a rotational parameter corresponding to a seed filter of the plurality of seed filters.

10. A system for determining optimized deep learning architecture, comprising:
a processor unit to:
receive a plurality of training images and a plurality of real time images corresponding to a subject;
receive a plurality of learning parameters comprising a plurality of filter classes and a plurality of architecture parameters;
determine a deep learning model based on the plurality of learning parameters and the plurality of training images, wherein the deep learning model comprises a plurality of filters, wherein the determining the deep learning model comprises reducing a number of the plurality of learning parameters of the deep learning model based at least on one or more constraints that determine a coefficient set that is applied to the plurality of filters, and wherein the coefficient set is determined based on at least a sparsity constraint and a convexity constraint applied to rows or columns of a parameter matrix;
determine a health condition of the subject based on the plurality of real time images using the deep learning model; and
provide the health condition of the subject to a medical practitioner.

11. The system of claim 10 , wherein the deep learning model comprises a convolution neural network, and wherein the plurality of filters comprise at least one reusable filter.

12. The system of claim 11 , wherein the processor unit is configured to:
receive a plurality of input feature maps, wherein each input feature map of the plurality of input feature map is a representation of the plurality of real time images; and
generate a plurality of output feature maps based on the convolution neural network, wherein each of the plurality of output feature map is a linear combination of filtered version of the plurality of input feature maps, wherein each of the plurality of output feature map is a representation of the input feature maps.

13. The system of claim 12 , wherein the convolution neural network comprises a plurality of seed filters, wherein dimensionality of the seed filters is based on a size of the plurality of training images.

14. The system of claim 13 , wherein the processor unit is configured to select a plurality of optimal filters among the plurality of seed filters for generating an output feature map of the plurality of output feature maps.

15. The system of claim 13 , wherein the processor unit is configured to determine the coefficient set corresponding to each output feature map of the plurality of output feature maps based on the plurality of seed filters.

16. The system of claim 13 , wherein the processor unit is configured to determine a dictionary of optimal filters for generating the plurality of output feature maps based on the plurality of seed filters.

17. The system of claim 16 , wherein the processor unit is further configured to determine the coefficient set corresponding to each output feature map of the plurality of output feature maps based on the dictionary of optimal filters.

18. The system of claim 13 , wherein the processor unit is further configured to determine a rotational parameter corresponding to a seed filter among the plurality of seed filters.

19. A health diagnostic system comprising:
an imaging modality;
a sub-system comprising:
a processor unit to:
receive a plurality of training images and a plurality of real time images corresponding to a subject;
receive a plurality of learning parameters comprising a plurality of filter classes and a plurality of architecture parameters;
determine a deep learning model based on the plurality of learning parameters and the plurality of training images, wherein the deep learning model comprises a plurality of filters, wherein at least one of the plurality of filters is a reusable filter, wherein the determining the deep learning model comprises determining a coefficient set that is applied to the plurality of filters, wherein the coefficient set is determined based on at least a sparsity constraint and a convexity constraint applied to rows or columns of a parameter matrix;
determine a health condition of the subject based on the plurality of real time images using the deep learning model; and
provide the health condition of the subject to a medical practitioner.

20. The health diagnost

Assignees

• Gen Electric

Inventors

• Thiruvenkadam Sheshadri
• Ranjan Sohan Rashmi
• Vaidya Vivek Prabhakar
• Ravishankar Hariharan
• Venkataramani Rahul
• Sudhakar Prasad

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title

• G06F18/24133

  Distances to prototypes · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title