Electronic apparatus and method for controlling thereof

Patent metadata

Field	Value
Publication number	US-2021271981-A1
Application number	US-202117171582-A
Country	US
Kind code	A1
Filing date	Feb 9, 2021
Priority date	Mar 2, 2020
Publication date	Sep 2, 2021
Grant date	—

Abstract

An electronic apparatus performing an operation of a neural network model is provided. The electronic apparatus includes a memory configured to store weight data including quantized weight values of the neural network model; and a processor configured to obtain operation data based on input data and binary data having at least one bit value different from each other, generate a lookup table by matching the operation data with the binary data, identify operation data corresponding to the weight data from the lookup table, and perform an operation of the neural network model based on the identified operation data.

First claim

What is claimed is:

1 . An electronic apparatus that performs an operation of a neural network model, the electronic apparatus comprising:
a memory configured to store weight data including quantized weight values of the neural network model; and a processor configured to:
obtain operation data based on input data and binary data having at least one bit value different from each other,
generate a lookup table by matching the operation data with the binary data,
identify operation data corresponding to the weight data from the lookup table, and
perform an operation of the neural network model based on the identified operation data.

2 . The electronic apparatus of claim 1 , wherein the binary data includes n bit values,
wherein the input data includes a plurality of input values of a matrix, and wherein the processor is further configured to:
obtain n input values in each column of the matrix, and
identify the operation data for each of the binary data based on the binary data and the n input values.

3 . The electronic apparatus of claim 2 , wherein the weight data includes a plurality of weight values of the matrix, and
wherein the processor is further configured to:
identify n weight values corresponding to the n input values in each row of the matrix,
identify binary data corresponding to the identified n weight values among the binary data, and
identify operation data corresponding to the identified binary data from the lookup table.

4 . The electronic apparatus of claim 3 , wherein the processor is further configured to:
determine, among a plurality of lookup tables generated based on input values of each column of the matrix, a lookup table corresponding to each column of an output matrix for the input data, and derive output values for each column of the output matrix from each of the lookup tables.

5 . The electronic apparatus of claim 3 , wherein the processor is further configured to:
divide the matrix including the plurality of input values into a first matrix and a second matrix based on predetermined rows, divide the matrix including the plurality of weight values into a third matrix and a fourth matrix based on predetermined columns, generate a first plurality of lookup tables based on input values of each column of the first matrix, identify first operation data corresponding to each row of the third matrix from the first plurality of lookup tables, generate a second plurality of lookup tables based on input values of each column of the second matrix, and identify second operation data corresponding to each row of the fourth matrix from the second plurality of lookup tables.

6 . The electronic apparatus of claim 2 , wherein the processor is further configured to:
obtain eight input values from each column of the matrix, and identify the operation data for each of the binary data based on the binary data and the eight input values.

7 . The electronic apparatus of claim 2 , wherein the processor is further configured to:
based on a first operation expression and a second operation expression having a same intermediate operation expression existing in a plurality of operation expressions based on the binary data and the n input values, perform the operation of the second operation expression based on the operation value of the first operation expression.

8 . A method for controlling an electronic apparatus to perform an operation of a neural network model, the method comprising:
obtaining operation data based on input data and binary data including at least one bit value different from each other; generating a lookup table by matching the operation data with the binary data; identify operation data corresponding to weight data including quantized weight values of the neural network model from the lookup table; and performing an operation of the neural network model based on the identified operation data.

9 . The method of claim 8 , wherein each of the binary data includes n bit values,
wherein the input data includes a plurality of input values of a matrix, and wherein the obtaining operation data comprises:
obtaining n input values in each column of the matrix; and
identifying the operation data for each of the binary data based on the binary data and the n input values.

10 . The method of claim 9 , wherein the weight data includes a plurality of weight values of a matrix, and
wherein performing the operation of the neural network model comprises:
identifying n weight values corresponding to the n input values in each row of the matrix,
identifying binary data corresponding to the identified n weight values among the binary data,
identifying the operation data corresponding to the identified binary data from the lookup table, and
performing the operation of the neural network model based on the identified operation data.

11 . The method of claim 10 , wherein performing the operation of the neural network model comprises:
determining, among a plurality of lookup tables generated based on input values of each column of the matrix, a lookup table corresponding to each column of an output matrix for the input data, and identifying output values of each column of the output matrix from each of the lookup tables.

12 . The method of claim 10 , wherein identifying the operation data comprises:
dividing the matrix including the plurality of input values into a first matrix and a second matrix based on predetermined rows; dividing the matrix including the plurality of weight values into a third matrix and a fourth matrix based on predetermined columns; generating a first plurality of lookup tables based on input values of each column of the first matrix; identifying first operation data corresponding to each row of the third matrix from the first plurality of lookup tables; generating a second plurality of lookup tables based on input values of each column of the second matrix; and identifying second operation data corresponding to each row of the fourth matrix from the second plurality of lookup tables.

13 . The method of claim 9 , wherein identifying the operation data comprises:
obtaining eight input values from each column of the matrix; and identifying the operation data for each of the binary data based on the binary data and the eight input values.

14 . The method of claim 9 , wherein generating the lookup table comprises:
based on a first operation expression and a second operation expression having a same intermediate operation expression existing in a plurality of operation expressions based on the binary data and the n input values, performing the operation of the second operation expression based on the operation value of the first operation expression.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Lee Dongsoo
• Park Baeseong
• Kim Byeoungwook
• Kwon Sejung
• Jeon Yongkweon

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06F1/03 Primary

  working, at least partly, by table look-up (G06F1/025 takes precedence) · CPC title

• G06N3/063 Primary

  using electronic means · CPC title

• G06N3/10 Primary

  Interfaces, programming languages or software development kits, e.g. for simulating neural networks · CPC title