Probabilistic Metaplasticity for Continual Learning with Memristors

Crossbar architectures utilizing memristor devices hold promise to address continual learning challenges in resource-constrained edge devices. However, these nanoscale devices often exhibit low precision and high variability in conductance modulation, rendering them unsuitable for continual learning solutions that consolidate weights through precise modulation. This issue can be circumvented by accumulating weight gradients in auxiliary high-precision memory and updating memristor weights when gradients are equivalent to memristor weight resolution. However, it leads to frequent memory access, high memory overhead, and energy dissipation. In this research, we propose probabilistic metaplasticity, which consolidates weights by modulating their update probability rather than magnitude. The proposed mechanism eliminates high-precision modification to weight magnitude and consequently, high-precision memory for gradient accumulation. We demonstrate the efficacy of the proposed mechanism by integrating probabilistic metaplasticity into a spiking network trained on an error threshold with low-precision memristor weights. Evaluations of two continual learning benchmarks show that probabilistic metaplasticity consumes ~67% lower memory for additional parameters and up to two orders of magnitude lower energy during parameter updates compared to an auxiliary memory-based solution while achieving state-of-the-art performance. The proposed model shows potential for energy-efficient continual learning with low-precision emerging devices.