Sequence Learning with Analog Neuromorphic Multi-Compartment Neurons and On-Chip Structural STDP

Neuromorphic computing is a candidate for advancing today’s AI systems towards fast and efficient online learning and inference. By exploiting biological principles, mixed-signal neuromorphic chips are well suited for emulating spiking neural networks (SNNs). Nevertheless, especially time-coded SNNs tend to struggle with the noise, uncertainty, and heterogeneity introduced by analog neuromorphic. We improve the robustness of the spiking hierarchical temporal memory (S-HTM) by removing its dependency on exact spike times and thereby enabling its deployment on the analog neuromorphic system BrainScaleS-2. Specifically, we introduce a new, adapted learning rule, implement it on-chip and evaluate it in a fully neuromorphic experiment using analog multi-compartment neurons and synapses on BrainScaleS-2 to learn sequences of symbols. Our results demonstrate that, while the on-chip network generates some overlapping predictions, potentially leading to contextual ambiguity, it is still capable of learning new sequences quickly and robustly, in some cases even faster than the original simulated S-HTM. We further show that the system’s natural heterogeneity, caused by its analog components, can replace the artificial heterogeneity introduced in the simulated network. Overall, the proposed network for BrainScaleS-2 can learn the presented sequences reliably without requiring exact spike times, demonstrating its increased robustness to noise caused by the system’s analog neurons and synapses.