Accelerating constraint-based neural network repairs by example prioritization and selection  

作　　者：Long ZHANG Shuo SUN Jun YAN Jian ZHANG Jiangzhao WU Jian LIU 

机构地区：[1]Institute of Software,Chinese Academy of Sciences,Beijing 100190,China [2]National Intelligent Voice Innovation Center,Hefei 231200,China [3]National China Electronic Product Reliability and Environmental Testing Research Institute,Guangzhou 510000,China

出　　处：《Frontiers of Computer Science》2025年第4期125-127,共3页计算机科学前沿(英文版)

基　　金：supported by the National Natural Science Foundation of China(Grant No.62132020);the Major Project of ISCAS(ISCAS-ZD-202302).

摘　　要：1 Introduction Compared with retraining,fine-tuning,and other traditional approaches,neural network repair approaches[1-7]can significantly improve the robustness of neural networks with lower time and computing cost.These repair techniques encode the anticipated performance of the repaired neural network into a Satisfiability Modulo Theories(SMT)problem and utilize an SMT solver to calculate a parameter matrix for the fully connected layer.This matrix can then be multiplied with the example feature vector to yield a vector that satisfies predetermined conditions.

关 键 词：REPAIRS constraint based parameter matrix neural networks repair techniques repaired neural network accelerating neural network 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]