Coding and decoding method based on block cyclic sparse matrix neural network

Abstract

The invention relates to a coding and decoding method based on a block cyclic sparse matrix neural network, and the method comprises the steps: compressing and coding a fully connected neural networkwith a block cyclic sparse weight matrix; using the cyclic property and sparsity of the weight matrix to design a coding method using a mask matrix and a non-zero value list; using the sparsity of theinput excitation to design a coding method using a mask vector and a non-zero value list. The method makes full use of the characteristics of the mask matrix and the cyclic matrix, and employs a hardware-friendly decoding method. The beneficial effects are that the method performs the simultaneous compressing and coding of a sparse input excitation vector and the weight matrix, effectively reduces the storage space required for the data and the number of storage accesses required to carry the data in the operation process. In the process of neural network operation, the method can effectivelyreduce the energy consumed by the memory access, brings convenience to a processor for skipping unnecessary operations, and improves the throughput of a hardware system.

Description

technical field [0001] The invention relates to a sparse deep neural network compression method, in particular to an encoding and decoding method based on a block cycle sparse matrix neural network. Background technique [0002] Deep neural networks are widely used in the field of artificial intelligence, especially in the field of image recognition, which has achieved the best accuracy so far. A large-scale deep neural network has high computational complexity and contains a large number of operational parameters, so it has high requirements for the operational performance of the processor. Processing deep neural networks in resource-constrained systems such as embedded systems requires high processor power efficiency. The study found that the deep neural network contains a large number of zero elements, and the sparsity of the neural network will be higher after pruning technology. Therefore, using the sparsity of deep neural networks to design special hardware accelerat...

AI Technical Summary

Problems solved by technology

The current accelerators for sparse neural networks have problems such as unbalanced load, complex encoding and decoding methods, and irregular operations.

Images