Physical reservoir computing / liquid state machine

Reservoir computing uses a fixed, randomly-connected recurrent dynamical system (the reservoir) to nonlinearly project inputs into a high-dimensional state, then trains only a simple linear readout. The spiking version is the liquid state machine. Because only the readout is trained, it is cheap to adapt and well-suited to hardware where the reservoir can be a physical substrate — a recurrent silicon-neuron network, a memristive network, or another nonlinear device — that computes for free.

For BCI this offers fast, low-cost adaptation of the readout to a nonstationary signal, but performance hinges on the reservoir having suitable dynamics (near the edge-of-chaos regime), and it generally trails end-to-end trained networks on the hardest decoding tasks. It remains a niche but active neuromorphic approach.