Adjustable hardware aware pruning and mapping framework based on ReRAM neural network accelerator

Abstract

The invention provides an adjustable hardware perception pruning and mapping framework based on a ReRAM neural network accelerator. The pruning and mapping framework comprises a DDPG agent and the ReRAM neural network accelerator. The DDPG agent is composed of a behavior decision module Actor and an evaluation module Critic, and the behavior decision module Actor is used for making a pruning decision for the neural network; the ReRAM neural network accelerator is used for mapping a model formed under a pruning decision generated by the behavior decision module Actor, and feeding back a performance parameter mapped by the model under the pruning decision to the evaluation module Critic as a signal; the performance parameters comprise energy consumption, delay and model accuracy of the simulator; the judgment module Critic updates a reward function value according to the fed back performance parameters and guides a pruning decision of the behavior decision module Actor in the next stage;according to the method, a pruning scheme which is most matched with hardware and user requirements and is most efficient is made by utilizing a reinforcement learning DDPG agent, so that the delay performance and the energy consumption performance on the hardware are improved while the accuracy is ensured.

Description

technical field [0001] The invention relates to the field of computer science and artificial intelligence, in particular to an adjustable hardware-aware pruning and mapping framework based on a ReRAM neural network accelerator. Background technique [0002] Deep neural networks play an important role in promoting the development of computer vision, natural language processing, robotics and other fields. With the development of mobile Internet of Things platforms, the application of neural networks on IoT devices is developing rapidly. Due to the computation-intensiveness and massive data mobility of neural networks, the application of neural networks will generate high energy consumption and high latency. However, IoT devices require more efficient neural network mapping due to limited computing resources and limited energy support on IoT platforms. solutions to reduce energy consumption and latency. Resistive Random Access Memory (ReRAM) due to its extremely low energy lea...

AI Technical Summary

Problems solved by technology

However, when the hardware specifications and types of ReRAM neural network accelerators are different, and when faced with different levels of user requirements such as delay and energy consumption, the traditional deep learning pruning scheme cannot perceive the changes in hardware and user requirements and make the same A pruning scheme, resulting in the performance inefficiency of model mapping on the ReRAM neural network accelerator hardware, which restricts the development of the performance advantages of the ReRAM neural network accelerator

Images