A deep neural network compression method and device

Abstract

A deep neural network compression method and device are proposed. The connection relationship between neurons of a neural network is usually represented by multiple matrices. The neural network compression method (900) according to the present invention includes: rearranging all rows of the plurality of matrices across matrices (S910), wherein the rearranged matrix rows are sequentially divided into multiple sub-matrices; Sensitivity analysis is performed on the sub-matrices to determine an initial compression ratio of the neural network (S920); the plurality of sub-matrices are compressed according to the determined initial compression ratio to obtain a compressed neural network (S930). The invention can ensure the load balancing of parallel computing processing units, so as to achieve the effects of releasing storage resources, accelerating computing speed and reducing power consumption.

Description

technical field [0001] The present invention relates to an artificial neural network, and more particularly to a deep neural network compression method and device. Background technique [0002] Artificial Neural Networks (ANN), also referred to as Neural Networks (NNs), is a mathematical computing model that imitates the behavioral characteristics of animal neural networks and performs distributed parallel information processing. In recent years, neural networks have developed rapidly and are widely used in many fields, such as image recognition, speech recognition, natural language processing, weather forecast, gene expression, content push and so on. [0003] In the neural network, there are a large number of connected nodes (also called "neurons"), and have two characteristics: 1) Each neuron, through a specific output function (also called activation function, activation function), calculates Process weighted input values ​​from other adjacent neurons; 2) The intensity ...

AI Technical Summary

Problems solved by technology

[0014] Embodiments of the present invention provide a deep neural network compression method and device, which are used to solve the defects of the prior art related to the neural network, such as large amount of calculation, decreased compression rate, and unsatisfactory compression effect, and realize load balancing of parallel computing processing units, thereby Achieve the effects of releasing storage resources, speeding up computing speed, and reducing power consumption, so that the performance of FPGA and other hardware implementation structures can be fully utilized

Images