Reducing computations in a neural network

Patent metadata

Field	Value
Publication number	US-10460230-B2
Application number	US-201615014686-A
Country	US
Kind code	B2
Filing date	Feb 3, 2016
Priority date	Jun 4, 2015
Publication date	Oct 29, 2019
Grant date	Oct 29, 2019

Abstract

Reducing computations in a neural network may include determining a group including a plurality of convolution kernels of a convolution stage of a neural network. The convolution kernels of the group are similar to one another. A base convolution kernel for the group may be determined. Scaling factors for a plurality of input feature maps processed by the group may be calculated. The convolution stage of the neural network may be modified to calculate a composite input feature map using the scaling factors and apply the base convolution kernel to the composite input feature map.

First claim

What is claimed is:

1. A method, comprising:
determining, using a processor, a group comprising a plurality of convolution kernels of a convolution stage of a neural network in which the convolution kernels of the group are similar to one another, the determining of the group by:
determining a similarity metric between a first convolution kernel and a second convolution kernel of the plurality of convolution kernels by calculating an element-wise ratio of the first convolution kernel to the second convolution kernel and calculating a standard deviation for the element-wise ratio of the first convolution kernel and the second convolution kernel; and
including the first convolution kernel and the second convolution kernel in the group responsive to determining that the similarity metric meets a similarity criterion;
determining a base convolution kernel for the group;
determining scaling factors based on the base convolution kernel for the group for a plurality of input feature maps processed by the group, each input feature map corresponding to a scaling factor and a convolution kernel; and
modifying the convolution stage of the neural network so that, when the neural network is executed, the convolution stage of the neural network performs:
scaling each of a plurality of input feature maps using the scaling factors,
generating a composite input feature map as a sum of the scaled input feature maps,
applying the base convolution kernel to the composite input feature map, and
summing the composite input feature map with any of a second plurality of input feature maps which are processed by a convolution kernel other than the base convolution kernel.

2. The method of claim 1 , wherein determining the group of convolution kernels further comprises:
selecting a convolution kernel from among a plurality of convolution kernels that operate on the feature map as a primary convolution kernel; and
calculating an element-wise ratio of the primary convolution kernel to other convolution kernels of the plurality of convolution kernels not yet assigned to a group.

3. The method of claim 1 , wherein determining the base convolution kernel for the group comprises:
selecting a convolution kernel of the group as the base convolution kernel according to a base selection metric.

4. The method of claim 1 , wherein determining the base convolution kernel for the group comprises:
determining the base convolution kernel as a function of the convolution kernels of the group.

5. The method of claim 1 , further comprising:
retraining the neural network comprising the modified convolution stage without permitting changes to membership in the group of convolution kernels.

6. The method of claim 1 , further comprising:
executing the neural network comprising the modified convolution stage.

7. A system, comprising:
a processor programmed to initiate executable operations comprising:
determining a group comprising a plurality of convolution kernels of a convolution stage of a neural network in which the convolution kernels of the group are similar to one another, the determining of the group by:
determining a similarity metric between a first convolution kernel and a second convolution kernel of the plurality of convolution kernels by calculating an element-wise ratio of the first convolution kernel to the second convolution kernel and calculating a standard deviation for the element-wise ratio of the first convolution kernel and the second convolution kernel; and
including the first convolution kernel and the second convolution kernel in the group responsive to determining that the similarity metric meets a similarity criterion;
determining a base convolution kernel for the group;
determining scaling factors based on the base convolution kernel for the group for a plurality of input feature maps processed by the group, each input feature map corresponding to a scaling factor and a convolution kernel; and
modifying the convolution stage of the neural network so that, when the neural network is executed, the convolution stage of the neural network performs:
scaling each of a plurality of input feature maps using the scaling factors,
generating a composite input feature map as a sum of the scaled input feature maps,
applying the base convolution kernel to the composite input feature map, and
summing the composite input feature map with any of a second plurality of input feature maps which are processed by a convolution kernel other than the base convolution kernel.

8. The system of claim 7 , wherein determining the group of convolution kernels further comprises:
selecting a convolution kernel from among a plurality of convolution kernels that operate on the feature map as a primary convolution kernel; and
calculating an element-wise ratio of the primary convolution kernel to other convolution kernels of the plurality of convolution kernels not yet assigned to a group.

9. The system of claim 7 , wherein determining the base convolution kernel for the group comprises:
selecting a convolution kernel of the group as the base convolution kernel according to a base selection metric.

10. The system of claim 7 , wherein determining the base convolution kernel for the group comprises:
determining the base convolution kernel as a function of the convolution kernels of the group.

11. The system of claim 7 , wherein the processor is further programmed to initiate executable operations comprising:
retraining the neural network comprising the modified convolution stage without permitting changes to membership in the group of convolution kernels.

12. A non-transitory computer-readable storage medium having instructions stored thereon which, when executed by a processor, perform a method comprising:
determining, using the processor, a group comprising a plurality of convolution kernels of a convolution stage of a neural network in which the convolution kernels of the group are similar to one another, the determining of the group by:
determining a similarity metric between a first convolution kernel and a second convolution kernel of the plurality of convolution kernels by calculating an element-wise ratio of the first convolution kernel to the second convolution kernel and calculating a standard deviation for the element-wise ratio of the first convolution kernel and the second convolution kernel; and
including the first convolution kernel and the second convolution kernel in the group responsive to determining that the similarity metric meets a similarity criterion;
determining a base convolution kernel for the group;
determining scaling factors based on the base convolution kernel for the group for a plurality of input feature maps processed by the group, each input feature map corresponding to a scaling factor and a convolution kernel; and
modifying the convolution stage of the neural network so that, when the neural network is executed, the convolution stage of the neural network performs:
scaling each of a plurality of input feature maps using the scaling factors,
generating a composite input feature map as a sum of the scaled input feature maps,
applying the base convolution kernel to the composite input feature map, and
summing the composite input feature map with any of a second plurality of input feature maps which are processed by a convolution kernel other than the base convolution kernel.

13. The non-transitory computer-readable storage medium of claim 12 , wherein determining the group of convolution kernels further comprises:
selecting a convolution kernel from among a plurality of convolution kernels that operate on the featu

Assignees

• Samsung Electronics Co Ltd

Inventors

• Brothers John W
• Lee Joohoon

Classifications

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06N3/082

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

• G06N3/0454 Primary

  Physics · mapped topic

• G06V10/82 Primary

  using neural networks · CPC title

• G06N3/0464 Primary

  Convolutional networks [CNN, ConvNet] · CPC title