US 10,592,799 B1Granted

Determining FL value by using weighted quantization loss values to thereby quantize CNN parameters and feature values to be used for optimizing hardware applicable to mobile devices or compact networks with high precision

AssigneeStradVision, Inc.

InventorsKye-Hyeon Kim, Yongjoong Kim, Insu Kim, Hak-Kyoung Kim, Woonhyun Nam, SukHoon Boo, Myungchul Sung, Donghun Yeo, Wooju Ryu, Taewoong Jang, Kyungjoong Jeong, Hongmo Je, Hojin Cho

Life of this patent

Estimating the term…

patentgrep

Loading patent…

Abstract

There is provided a method for determining an FL value to be used for optimizing hardware applicable to mobile devices, compact networks, and the like with high precision. The method includes steps of: a computing device (a) applying quantization operations to original values included in an original vector by referring to a BW value and each of FL candidate values, to thereby generate each of quantized vectors, including the quantized values, corresponding to each of the FL candidate values; (b) generating each of weighted quantization loss values, corresponding to each of the FL candidate values, by applying weighted quantization loss operations to information on each of differences between the original values and the quantized values included in each of the quantized vectors; and (c) determining the FL value among the FL candidate values by referring to the weighted quantization loss values and a device using the same.

Claims16

Dependent claims

A method for determining an FL value, which is an FL parameter that is an absolute value of an exponent of a number represented by a LSB, least significant bit, included in quantized values, comprising steps of:

(a) a computing device, if an original vector including original values is acquired, applying quantization operations to the original values by referring to (i) a BW, bit width, value, which is a BW parameter corresponding to the number of bits used to represent each of the quantized values, and (ii) each of FL candidate values, to thereby generate each of quantized vectors, including the quantized values, corresponding to each of the FL candidate values;
b) the computing device generating each of weighted quantization loss values, corresponding to each of the FL candidate values, by applying weighted quantization loss operations to information on each of differences between the original values included in the original vector and the quantized values included in each of the quantized vectors;
and
(c) the computing device determining the FL value among the FL candidate values by referring to the weighted quantization loss values, wherein
at the step of (b), the weighted quantization loss operation follows a formula below,
Ll=∥{right arrow over (W)}·({right arrow over (Q)}l−{right arrow over (x)})∥p
wherein {right arrow over (x)} is the original vector, Ql is an l-th quantized vector, {right arrow over (W)} is a weight vector, which includes at least part of weight values generated by referring to sizes of the original values · is an operation that generates an integrated vector whose elements are determined as results of element-wise multiplication of the weight vector {right arrow over (W)} and a difference vector ({right arrow over (Q)}l−{right arrow over (x)}), and ∥ ∥p is an L-P norm operation.

4 more elements

A computing device for determining an FL value, which is an FL parameter that is an absolute value of an exponent of a number represented by a LSB, least significant bit, included in quantized values, comprising:

at least one memory that stores instructions;
and
at least one processor configured to execute the instructions to:
perform processes of (I) applying quantization operations to the original values by referring to (i) a BW, bit width, value, which is a BW parameter corresponding to the number of bits used to represent each of the quantized values, and (ii) each of FL candidate values, to thereby generate each of quantized vectors, including the quantized values, corresponding to each of the FL candidate values;
(II) generating each of weighted quantization loss values, corresponding to each of the FL candidate values, by applying weighted quantization loss operations to information on each of differences between the original values included in an original vector including original values and the quantized values included in each of the quantized vectors;
and (III) determining the FL value among the FL candidate values by referring to the weighted quantization loss values, wherein
at the process of (II), the weighted quantization loss operation follows a formula below,
Ll=∥{right arrow over (W)}·({right arrow over (Q)}l−{right arrow over (x)})∥p
wherein {right arrow over (x)} is the original vector, Ql is an l-th quantized vector, {right arrow over (W)} is a weight vector, which includes at least part of weight values generated by referring to sizes of the original values · is an operation that generates an integrated vector whose elements are determined as results of element-wise multiplication of the weight vector {right arrow over (W)} and a difference vector ({right arrow over (Q)}l−{right arrow over (x)}), and ∥ ∥p is an L-P norm operation.

6 more elements

Dependencies are inferred from the claim text.

Facts

Application: 16/255,012Filed 23 Jan 2019
Granted: 17 Mar 2020
Term ends: Not estimatedThe source records don’t support an estimate.

Classification7 codes

G06N 3/0495
Architecture, e.g. interconnection topology: quantised networks; Sparse networks; Compressed networks
G06N 3/063
Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons: using electronic means
G06F 17/16
Complex mathematical operations: matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
G06N 3/045
Architecture, e.g. interconnection topology: combinations of networks