Neural network model block compression method, training method, computing device and system

Abstract

A network model block compression method for a neural network, comprising: a weight matrix obtaining step, obtaining a weight matrix of a trained network model of a neural network; a weight matrix blocking step, dividing the weight matrix into an array consisting of a plurality of initial sub-blocks according a predetermined array size; a step of concentrating the weighted elementto be cut, according to the absolute values and values of the weights of the matrix elements in the sub-blocks, concentrating the matrix elements with smaller weights to the sub-blocks to be cut through row and columnexchange, so that the absolute value and value of the weights of the matrix elements in the sub-blocks to be cut are made smaller relative to the absolute values and values of the weights of the matrix elements in other sub-blocks that are not in the sub-blocks to be cut; a sub-block cut step, cutting the weights of the matrix elements in the above-mentioned sub-blocks to be cut, and obtaining the final weight matrix to implement compression of the network model of the neural network. The resource overhead can be saved, and a large-scale neural network can be arranged under the condition of limited resources.

Description

technical field [0001] The present invention generally relates to the technical field of neural networks, and more specifically relates to a network model block compression method, a training method, a computing device and a hardware system for neural networks. Background technique [0002] With the gradual failure of Moore's Law, the progress of traditional chip technology has slowed down, and people have to face new applications and new devices. In recent years, neural network (Neural Network, NN) computing has made breakthroughs, and has achieved high accuracy in many fields such as image recognition, language recognition, and natural language processing. However, neural networks require massive computing resources. Traditional It is difficult for general-purpose processors to meet the computing needs of deep learning, and designing dedicated chips has become an important development direction. [0003] Specifically, the modeling of neural networks is usually constructed...

AI Technical Summary

Problems solved by technology

[0019] Although these existing technologies can greatly reduce the size of the model, they are not suitable for the application of chips based on memristors and TrueNorth that can perform matrix-vector multiplication operations. neural network model

For example...

Images