BBS-RNN

Abstract

Large deep neural network (RNN) models are a key challenge to energy efficiency because out-of-chip DRAM access is much more energy intensive than arithmetic or SRAM operations. It uses three main methods to promote the in-depth study of model compression. Block-circulant com- pression method adopts the structure of block-based circulant matrix to realize weight sharing and fast Fourier matrix operation; Weight pruning takes advantage of the redundancy in weight quantity and can be carried out in the case of unstructured way and structured way. For the unstructured way, higher flexibility and pruning rate can be achieved while, due to the irregularity of weight, index access is increased, and for the structured way, the complete matrix structure can be kept at a lower pruning rate; Weight quantization takes advantage of the redundancy of the median weight. Quantification is more hardware-friendly than pruning, and has become a "must-do" step for FPGA and ASIC implementations. Recursive neural network (RNNs) is becoming more and more important in the applica- tion of time series correlation, which needs efficient and real-time implementation. The two main types are long-short term memory (LSTM) and gated recursive unit (GRU) networks. Real-time, efficient and accurate hardware RNN implementation is a challenging task because it is highly sen- sitive to error accumulation and requires special activation function implementation. This paper presents a full-stack RNN framework based on block-based structure prun- ing(BSP). BSP RNN model has good pruning granularity, which is helpful to improve pruning rate and has regular structure, which is beneficial to hardware parallelism. On this basis, we compare different compression methods and discuss the unique characteristics of different RNNs. And we provide constructive suggestions for the future research of RNN compression.--Author's abstract