MONO: Enhancing Bit-Flip Resilience With Bit Homogeneity for Neural Networks

Abstract

Deep neural networks (DNNs) have been applied across diverse domains, including safety-critical applications. Past studies indicate that DNNs are very sensitive to changes in weights and activations due to uneven bit-weight distribution in standard number formats like fixed points, which can cause significant output accuracy fluctuations. To address this issue, we introduce a new data type called MONO to enhance bit-flip resilience using uniformity at the bit level by employing symmetric weights for all bit positions. On average, MONO has improved error resilience more effectively than the fixed-point data type, even when utilizing triple modular redundancy (TMR) and most significant bit (MSB) protection, while maintaining low overhead.