Convolution neural network training apparatus and method thereof

Patent metadata

Field	Value
Publication number	US-2016162782-A1
Application number	US-201514960942-A
Country	US
Kind code	A1
Filing date	Dec 7, 2015
Priority date	Dec 9, 2014
Publication date	Jun 9, 2016
Grant date	—

Abstract

An apparatus and method of training a convolutional neural network (CNN) are provided. A method of training a CNN including a plurality of convolution layers stored in a memory involves approximating, using a processor, a convolution layer among the plurality of convolution layers using a low-rank approximation; reducing the number of output reconstruction filters of the approximated convolution layer; and modifying a structure of the CNN based on an approximation result and the reduced number of output reconstruction filters.

First claim

What is claimed is:

1 . A method of training a convolutional neural network (CNN) comprising a plurality of convolution layers stored in a non-transitory memory, the method comprising:
approximating, using a processor, a convolution layer among the plurality of convolution layers using a low-rank approximation; reducing the number of output reconstruction filters of the approximated convolution layer; modifying a structure of the CNN based on an approximation result and the reduced number of output reconstruction filters; and training the modified CNN.

2 . The method of claim 1 , after the training of the modified CNN, further comprising:
performing the operations of:
sequentially approximating convolution layers that follow the approximated convolution layer;
reducing the number of output reconstruction filters of the currently approximated convolution layer and modifying the structure of the CNN; and
training the modified CNN.

3 . The method of claim 1 , after the training of the modified CNN, further comprising:
classifying image data using the trained CNN; and in response to an accuracy of the classification not satisfying a designated criteria, performing the operations of:
reducing the number of output reconstruction filters;
modifying the structure of the CNN; and
training the modified CNN.

4 . The method of claim 1 , further comprising:
modifying a structure of at least one convolution layer that follows the approximated convolution layer, wherein the structure of the CNN is modified based on a result of the modifying of the structure of said at least one convolution layer.

5 . The method of claim 4 , wherein the structure of the CNN is modified by changing the number of convolution filters of the at least one convolution layer that follows the approximated convolution layer.

6 . The method of claim 4 , after the training of the modified CNN, further comprising:
classifying image data using the trained CNN; and in response to an accuracy of the classification not satisfying a designated criteria, performing the operations of:
modifying the structure of at least one convolution layer that follows the approximated convolution layer;
modifying the structure of the CNN; and
training the modified CNN.

7 . The method of claim 1 , wherein one convolution layer among the plurality of convolution layers is approximated into one or more input conversion filters, one or more convolution filters, and one or more output reconstruction filters.

8 . The method of claim 7 , wherein the input conversion filter is configured to reduce the number of channels of input data, the convolution filter is configured to perform a convolution operation on input data having a reduced number of channels, and the output reconstruction filter is configured to restore a convolution operation result to have the same number of channels as the number of channels of the convolution layer.

9 . A non-transitory computer-readable medium storing instructions that, when executed by a computer processor, cause the computer processor to train a convolution neural network stored in a non-transitory memory according to the method of claim 1 .

10 . An apparatus for training a convolution neural network (CNN) comprising a plurality of convolution layers stored in a non-transitory memory, the apparatus comprising:
an approximation processor configured to approximate a convolution layer among the plurality of convolution layers using a low-rank approximation; a filter count changer configured to reduce the number of output reconstruction filters of the approximated convolution layer; and a training processor configured to modify a structure of the CNN based on an approximation result and the reduced number of output reconstruction filters and to train the modified CNN.

11 . The apparatus of claim 10 , wherein the approximation processor is configured to sequentially approximate another convolution layer that follows the approximated convolution layer in response to the modified CNN being trained.

12 . The apparatus of claim 10 , further comprising:
a classifier configured to classify image data using the trained CNN, wherein the filter count changer re-changes the number of output reconstruction filters in response to an accuracy of the classification not satisfying a designated criteria.

13 . The apparatus of claim 10 , further comprising:
a layer structure modifier configured to modify a structure of at least one convolution layer that follows the approximated convolution layer, wherein the training processor modifies the structure of the CNN based on a result of modifying the structure of said convolution layer that follows the approximated convolution layer.

14 . The apparatus of claim 13 , wherein the layer structure modifier is configured to modify the structure of at least one convolution layer by changing the number of convolution filters of said convolution layer.

15 . The apparatus of claim 13 , further comprising:
a classifier configured to classify image data using the trained CNN, wherein the layer structure modifier is configured to re-modify the structure of the at least one convolution layer that follows the approximated convolution layer in response to an accuracy of the classification not satisfying a designated criteria.

16 . The apparatus of claim 10 , wherein the approximation processor is configured to approximate the convolution layer among the plurality of convolution layers into one or more input conversion filters, one or more convolution filters, and one or more output reconstruction filters.

17 . The apparatus of claim 16 , wherein the input conversion filter is configured to reduce the number of channels of input data, the convolution filter is configured to perform a convolution operation on input data having a reduced number of channels, and the output reconstruction filter is configured to restore a convolution operation result to have the same number of channels as the number of channels of the convolution layer.

18 . An apparatus for training a neural network, the apparatus comprising:
a non-transitory memory storing a convolution neural network (CNN) comprising a plurality of convolution layers; and a processor configured to approximate a convolution layer among the plurality of convolution layers using a low-rank approximation, reduce a number of output reconstruction filters of the approximated convolution layer, modify a structure of the CNN based on a result of the approximating of the convolution layer and the reduced number of output reconstruction filters, and train the modified CNN stored in the non-transitory memory.

19 . The apparatus of claim 18 , wherein the apparatus trains the modified CNN by retrieving input image data from a training data memory storage.

20 . The apparatus of claim 18 , wherein the processor is further configured to classify input image data retrieved from a training data memory storage by using the modified CNN, and further modify the modified CNN in response to an accuracy of the classification not satisfying a designated criteria.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Park Hyoung Min

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/082 Primary

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

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title