Analog Device Noise Boosts Continual Learning

In analog neuromorphic hardware, intrinsic device noise is typically seen as a detriment to accuracy. This research explores a counter-intuitive idea: can this noise actually be used to consolidate memories in continual learning? The authors introduce a concept called "Intrinsic-Noise Consolidation." The method reinterprets per-synapse memory consolidation as a Doob h-transform, where each weight's stochastic dynamics are conditioned on never crossing a critical barrier around its consolidated value. This conditioning introduces an extra drift term, a restoring force that is amplified by the noise variance itself and diverges at the barrier. Crucially, the research predicts and experimentally verifies a non-monotonic, inverted-U relationship between increasing intrinsic noise and sequential-task retention. This means there's an optimal level of noise that significantly improves memory retention, a phenomenon not observed in traditional anchored-drift methods. Experiments on both Split-MNIST and real BrainScaleS-2 silicon confirm that barrier-conditioning improves prior task retention, effectively turning analog noise into a valuable resource for stability-plasticity trade-offs.