Method for processing input on basis of neural network learning and apparatus therefor

Patent metadata

Field	Value
Publication number	US-2019251396-A1
Application number	US-201716338579-A
Country	US
Kind code	A1
Filing date	Oct 20, 2017
Priority date	Nov 7, 2016
Publication date	Aug 15, 2019
Grant date	—

Abstract

An electronic device is provided. The electronic device includes a memory storing parameter sets, each of which includes one weight and bias sets respectively corresponding to n (where, n>1) occlusion levels among a plurality of occlusion levels within a certain range and at least one processor configured to obtain output data by inputting input data to a neural network.

First claim

1 . An electronic device for processing an image based on neural network training, the electronic device comprising:
a memory storing a plurality of parameter sets, each of which includes one weight and bias sets respectively corresponding to n (where n>1) occlusion levels among a plurality of occlusion levels within a certain range; and at least one processor configured to obtain output data by inputting input data to a neural network, wherein the at least one processor is configured to:
obtain an occlusion value and determine a specific parameter set among the plurality of parameter sets based on the occlusion value,
determine a specific bias set in the specific parameter set based on the occlusion value, and
obtain the output data using a weight corresponding to the specific parameter set and the specific bias set.

2 . The electronic device of claim 1 ,
wherein the plurality of parameter sets include a first parameter set and a second parameter set, and wherein the first parameter set includes a plurality of first bias sets obtained for each occlusion level with respect to a first weight and the second parameter set includes a plurality of second bias sets obtained for each occlusion level with respect to a second weight.

3 . The electronic device of claim 2 ,
wherein the first weight is a value optimized for the neural network with respect to a minimum occlusion level in a range of the occlusion level, and wherein the second weight is a value optimized for the neural network with respect to a maximum occlusion level in the range of the occlusion level.

4 . The electronic device of claim 2 , wherein the at least one processor is configured to:
select the first parameter set, when the occlusion value meets a specific value, and select the second parameter set, when the occlusion value does not meet the specific value.

5 . The electronic device of claim 1 , wherein the at least one processor is configured to:
when one of the n occlusion levels is identical to the obtained occlusion value, determine the specific bias set as a bias set corresponding to the one of the n occlusion levels, and when the n occlusion levels are not identical to the obtained occlusion value, determine the specific bias set based on two or more of the n occlusion levels.

6 . The electronic device of claim 5 , wherein the at least one processor is configured to:
determine the specific bias set based on an interpolation of bias sets corresponding to the two or more of the n occlusion levels.

7 . The electronic device of claim 5 , wherein the at least one processor is configured to:
determine the specific bias set based on an extrapolation of bias sets obtained for the two or more of the n occlusion levels.

8 . The electronic device of claim 1 , wherein the at least one processor is configured to:
determine a bias set, corresponding to an occlusion level which is shortest from the obtained occlusion value among occlusion levels in the specific parameter set, as the specific bias set.

9 . The electronic device of claim 1 , wherein the neural network is a convolutional neural network.

10 . The electronic device of claim 1 , wherein the at least one processor is configured to:
generate a feature map of a convolution layer based on a weight corresponding to the specific parameter set and the specific bias set.

11 . The electronic device of claim 1 , wherein the bias set corresponds to a distance from a hyperplane for recognizing the input data to the origin.

12 . The electronic device of claim 1 , wherein the bias set includes a bias value associated with each convolution layer.

13 . The electronic device of claim 1 , wherein the bias set includes a bias value associated with each feature map of a convolution layer.

14 . The electronic device of claim 1 , wherein the at least one processor includes a graphic processing unit (GPU).

15 . An electronic device for processing an image based on neural network training, the electronic device comprising:
a memory storing a first parameter set and a second parameter set, each of which includes one weight and bias values respectively corresponding to n (where, n>1) occlusion levels among a plurality of occlusion levels within a certain range; and at least one processor configured to obtain output data by inputting input data to a neural network, wherein the at least one processor is configured to:
obtain an occlusion value and determine a parameter set to be applied to the neural network as the first parameter set based on the occlusion value,
determine a specific bias value in the first parameter set based on the occlusion value, and
obtain the output data using a weight corresponding to the first parameter set and the specific bias value.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Geraci James Russell
• Kapoor Parichay

Classifications

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

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

• G06F18/214

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

• G06N3/063

  using electronic means · CPC title