EPSILON: Adaptive Fault Mitigation in Approximate Deep Neural Network using Statistical Signatures

2025-04-24

AI Summary

Permanent hardware faults in approximate deep neural network accelerators (AxDNNs) cause severe accuracy degradation, while conventional fault-tolerance techniques incur prohibitive overhead and are ill-suited for energy-constrained edge deployments. To address this, we propose EPSILON—a lightweight, adaptive fault-tolerance framework. Its core innovations include: (i) a novel non-parametric pattern-matching algorithm based on statistical signatures, enabling constant-time, zero-interrupt fault detection; and (ii) a dynamic, layer-aware mitigation strategy selection mechanism guided by layer importance scoring and weight-distribution sensitivity analysis. Evaluated across datasets from MNIST to ImageNet-1k, EPSILON maintains 80.05% Top-1 accuracy while achieving 22% higher inference throughput and 28% improved energy efficiency over baseline AxDNNs. By jointly optimizing robustness, real-time performance, and energy efficiency, EPSILON is particularly well-suited for safety-critical edge computing applications.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

The increasing adoption of approximate computing in deep neural network accelerators (AxDNNs) promises significant energy efficiency gains. However, permanent faults in AxDNNs can severely degrade their performance compared to their accurate counterparts (AccDNNs). Traditional fault detection and mitigation approaches, while effective for AccDNNs, introduce substantial overhead and latency, making them impractical for energy-constrained real-time deployment. To address this, we introduce EPSILON, a lightweight framework that leverages pre-computed statistical signatures and layer-wise importance metrics for efficient fault detection and mitigation in AxDNNs. Our framework introduces a novel non-parametric pattern-matching algorithm that enables constant-time fault detection without interrupting normal execution while dynamically adapting to different network architectures and fault patterns. EPSILON maintains model accuracy by intelligently adjusting mitigation strategies based on a statistical analysis of weight distribution and layer criticality while preserving the energy benefits of approximate computing. Extensive evaluations across various approximate multipliers, AxDNN architectures, popular datasets (MNIST, CIFAR-10, CIFAR-100, ImageNet-1k), and fault scenarios demonstrate that EPSILON maintains 80.05% accuracy while offering 22% improvement in inference time and 28% improvement in energy efficiency, establishing EPSILON as a practical solution for deploying reliable AxDNNs in safety-critical edge applications.

Problem

• Mitigate permanent faults in approximate DNN accelerators efficiently
• Reduce overhead and latency in fault detection for AxDNNs
• Maintain accuracy while preserving energy benefits in AxDNNs

Innovation

• Lightweight framework using statistical signatures
• Non-parametric pattern-matching for fault detection
• Dynamic mitigation based on weight distribution analysis