SALMON: Self-Adaptive Learning Model on Neuromorphic Hardware

Analog in-memory computing (AIMC) is a promising technology for energy-efficient acceleration of deep learning workloads. While significant advancements have been achieved in accelerating on-chip inference, on-chip training has not received as much attention due to the challenges posed by the nonideal characteristics of AIMC. Self-adaptive learning model on neuromorphic hardware (SALMON) is introduced as a novel on-chip training method designed to operate effectively on large-scale neuromorphic computing systems. SALMON leverages a self-adaptive network (SAnet) that integrates an analog backbone network with attachable digital attention blocks to enhance network performance, indirectly addressing hardware nonidealities. The effectiveness of SALMON is demonstrated in CIFAR-10 image classification using the ResNet architecture with three distinct levels of analog device models. Performance trends in test accuracy reach 91.49%, reflecting a 13.1% improvement across network scales with varying nonideality levels. Ablation studies assess the impact and role of digital attention blocks based on Grad-CAM analysis. Finally, application strategies are presented to minimize the use of attachable digital components during on-chip training, significantly reducing relative power consumption by approximately 70% under hardware conditions.