Neural network configuration parameter training and deploying method and device for coping with device mismatch

Abstract

The invention discloses a neural network configuration parameter training and deployment method and device. The method comprises the steps of searching simulated attacked neural network configuration parameters based on neural network configuration parameters, so that the attacked neural network configuration parameters move in the direction deviating from the output result of a neural network corresponding to the neural network configuration parameters to the maximum extent; taking the difference between the current neural network configuration parameter and the neural network output result corresponding to the attacked neural network configuration parameter as a robust loss function, and taking the robust loss function as a part of a total loss function; and finally, optimizing neural network configuration parameters based on the total loss function. Especially for ultra-low power consumption sub-threshold and mixed signal circuits, the scheme can solve the problem of configuration parameter disturbance caused by device mismatch, and the technical effect of low-cost and high-efficiency neural network accelerator parameter deployment is achieved.

Description

technical field [0001] The invention relates to the field of training of neural network configuration parameters, in particular to the field of neural network training based on configuration parameters against attacks that can deal with the problem of device mismatch. Background technique [0002] Mixed-signal (mixed-signal) circuits have a very obvious power consumption advantage over digital circuits. Neural Network (NN) accelerators (hereinafter referred to as NN) accelerators may be implemented as mixed-signal circuits through silicon chip manufacturing methods. These fabrication methods introduce variables in the thickness and properties of the material layers from which the chip is made. These variables can cause the electrical behavior of fabricated transistors and other devices to vary widely across the surface of a single chip and vary from chip to chip, a phenomenon known as "device mismatch." Neuromorphic (also known as brain-like) neural network accelerators, s...

AI Technical Summary

Problems solved by technology

[0015] In view of the above-mentioned defects in the prior art, the technical problem to be solved by the present invention is: how to deal with the problem that the neural network task processing performance is weakened due to the comprehensive random disturbance of the configuration parameters of the neural network accelerator, and then the configuration parameters of the chip caused cannot Problems with Large-Scale Commercial Deployment

Images