NEURAL NETWORK DEVICES AND METHODS OF OPERATING THE SAME

• US 20180253635A1
• Filed: 01/08/2018
• Published: 09/06/2018
• Est. Priority Date: 03/03/2017
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: PARK, Jun-seok
• Granted Patent: US 11,295,195 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 2207/4824 Neural networks
• CPC Class Codes: G06F 7/523 Multiplying only
• CPC Class Codes: G06F 7/535 Dividing only
• CPC Class Codes: G06F 7/5443 Sum of products for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A neural network device may generate an input feature list based on an input feature map, where the input feature list includes an input feature index and an input feature value, generating an output feature index based on the input feature index corresponding to an input feature included in the input feature list and a weight index corresponding to a weight included in a weight list, and generating an output feature value corresponding to the output feature index based on the input feature value corresponding to the input feature and a weight value corresponding to the weight.

Claims

1. A method of operating a neural network device, the method comprising:

generating an input feature list based on an input feature map, the input feature list including an input feature index and an input feature value, the input feature index and the input feature value corresponding to an input feature;

generating an output feature index based on a first operation on the input feature index and a weight index of a weight list; and

generating an output feature value corresponding to the output feature index based on a second operation on the input feature value and a weight value corresponding to the weight index.

2. The method of claim 1, wherein the generating the input feature list includes generating the input feature list based on at least one input feature of a plurality of input features of the input feature map having a non-zero value.

3. The method of claim 1, wherein the weight list includes at least one weight index and at least one weight value, the at least one weight index and the at least one weight value corresponding to at least one weight of a plurality of weights of a weight map, the at least one weight having a non-zero value.

4. The method of claim 1, wherein the generating of the output feature value includes multiplying the input feature value by the weight value.

5. The method of claim 1, wherein the generating of the output feature value includes,

generating a multiplication value based on multiplying the input feature value by the weight value, the multiplication value corresponding to the output feature index;

and

generating the output feature value based on adding a plurality of multiplication values, the plurality of multiplication values corresponding to the output feature index.

6. The method of claim 1, wherein the generating the output feature index includes adding the input feature index and the weight index.

7. The method of claim 6, wherein the generating the output feature index further includes,

dividing an added-up value resulting from the adding by an integer;

and

selecting a quotient of the dividing as the output feature index based on a determination that no remainder is present upon completion of the dividing.

9. The method of claim 1, wherein the generating of the input feature list includes

generating an initial input feature list, the initial input feature list including an initial input feature index corresponding to a location of the input feature and the input feature value corresponding to the input feature, and

generating the input feature index padded with zeros based on adding a feature bias index to the initial input feature index.

10. The method of claim 1, further comprising generating the weight list from a weight map.

11. The method of claim 10, wherein the generating the weight list includes, generating an initial weight list, the initial weight list including an initial weight index corresponding to a location of a weight and a weight value of the weight from the weight map, and

adjusting the initial weight index to correspond to a particular operation.

12. The method of claim 11, wherein the adjusting the initial weight index includes,

mirroring the initial weight index based on a weight bias index indicating a center of a matrix of the weight map, and

subtracting the weight bias index from a mirrored weight index.

13. The method of claim 1, further comprising generating an information signal based on the output feature value.

8. (canceled)

14. A method of operating a neural network device, the method comprising:

generating an input feature list, the input feature list including an input feature index and an input feature value corresponding to an input feature having a non-zero value, the input feature index indicating a location of the input feature on an input feature map;

generating an output feature index based on an index operation on the input feature index; and

generating an output feature value corresponding to the output feature index based on a data operation on the input feature value.

15. The method of claim 14, wherein the generating the output feature index includes adding the input feature index and a weight index of a weight list, the weight index corresponding to a weight having a non-zero value.

16. The method of claim 15, wherein the generating the output feature value includes multiplying the input feature value by a weight value corresponding to the weight index.

17. The method of claim 14, wherein the generating of the output feature index includes

performing division on the input feature index based on a particular sub-sampling size, and

selecting a quotient of the division as the output feature index, the output feature index corresponding to the input feature.

18. The method of claim 14, wherein generating of the output feature value includes calculating the output feature value corresponding to the output feature index as

a maximum value among a plurality of input feature values corresponding to the output feature index, or

an average of the plurality of input feature values.

19-31. -31. (canceled)

32. A method comprising:

generating, using a list maker of a processor, an input feature list based on an input feature map, the input feature list including an input feature index and an input feature value, the input feature index and the input feature value corresponding to an input feature; and

causing an index remapper of the processor to perform a first operation to generate an output feature index, the first operation including

adding the input feature index and a weight index of a weight list,

dividing an added-up value resulting from the adding by an integer, and

selecting a quotient of the dividing as the output feature index based on a determination that no remainder is present upon completion of the dividing.

33. The method of claim 32, further comprising:

cause a data operation circuit to perform a second operation on the input feature value and a weight value corresponding to the weight index to generate an output feature value corresponding to the output feature index.

34. The method of claim 33, wherein the generating of the output feature value includes,

generating a multiplication value based on multiplying the input feature value by the weight value, the multiplication value corresponding to the output feature index;

and

generating the output feature value based on adding a plurality of multiplication values, the plurality of multiplication values corresponding to the output feature index.