On-chip learning (local plasticity)

On-chip learning updates the decoder's weights in the device itself, ideally with local rules that need only signals available at each synapse — spike-timing-dependent plasticity (STDP), local surrogate rules, or three-factor rules gated by a global reward/error signal — rather than full backpropagation, which is costly to implement in hardware. Memristive synapses that change conductance with pulses are a natural substrate. For BCI this promises continual, on-device adaptation to representational drift and to the user, without exporting data or a cloud retraining loop.

The reality is that purely local rules generally underperform gradient training on hard decoding tasks, and stable long-horizon on-chip learning must contend with device non-idealities and catastrophic forgetting. Most systems therefore train offline and deploy fixed weights, or do lightweight on-device recalibration, with fully autonomous on-chip learning still largely a research goal.