A Convolutional Neural Network Reinforcement Framework Based on Neural Network Interpretability
Year: 2025
Type: conference-paper
Abstract: With the wide deployment of deep learning models in various applications, the reliability problem of neural networks has gradually become a research hotspot. Computer systems are susceptible to single-particle effects, electromagnetic interference, and other factors in complex environments, resulting in frequent bit-flipping faults. Among them, Silent Data Corruption (SDC), as an imperceptible fault, cannot be detected by the system, which ultimately destroys the output results of neural network models. Current SDC detection techniques mainly focus on the system instruction stream level, and lack efficient methods to implement detection reinforcement at the model level. In addition, due to the black-box weight update mechanism of neural networks, traditional selective redundancy methods cannot provide a reliable explanation of which important weights to fix. To address these issues, this paper proposes a Convolutional Neural Network Reinforcement Framework Based on Neural Network Interpretability (RFNI). RFNI explores the redundancy ratio of convolutional kernels by performing layer-by-layer search in convolutional neural networks through SAC (Soft Actor-Critic) agent; secondly, the importance of each convolutional kernel in the current layer is calculated by combining with neural network interpretability algorithms; and finally, the importance of each convolutional kernel in the current layer is calculated based on the total Floating Point Operations Per Second (FLOPS). Second, the importance of each convolutional kernel in the current layer is calculated by combining with the neural network interpretable algorithm; finally, the granularity configurable redundancy reinforcement is implemented based on the total floating point operations per second (FLOPS). The experimental results show that RFNI on multiple neural network architectures significantly improves the recognition accuracy in case of SDC over existing methods, and can effectively enhance the robustness of the model.
Authors: Haocheng Song, Pengfei Yu, Jiayi Ou, Wei Li, Jingjing Gu
Institution: Nanjing University of Aeronautics and Astronautics
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Award: 面向城市演变多重关系驱动的群体出行流量预测的研究
OpenAlex ID: W4411015856