The leaky integrate-and-fire neuron
A spiking neural network decoder computes not with continuous activations but with discrete spikes in time, exactly like the biology it reads. Its workhorse unit is the leaky integrate-and-fire (LIF) neuron: it integrates incoming current, leaks toward rest, and emits a spike when its membrane potential crosses a threshold, then resets.
The LIF dynamics: membrane time constant au_m sets how fast the neuron forgets, input current I(t) drives it up, and a threshold-and-reset turns it into an event emitter. Stacking such units with weighted spike connections gives a decoder that is sparse and asynchronous by construction.
Training through the non-differentiable spike
Spikes are all-or-none, so the spike function has a derivative that is zero almost everywhere and infinite at threshold — ordinary backpropagation cannot flow through it. Two routes get around this. Surrogate-gradient training keeps the hard spike in the forward pass but substitutes a smooth, well-behaved derivative in the backward pass, letting standard deep-learning tooling train the network directly.
A typical surrogate gradient: a smooth bump centred on the threshold that stands in for the spike's true (degenerate) derivative during backprop. The sharpness $eta$ trades gradient locality against training stability.
The second route is ANN-to-SNN conversion: train an ordinary rate-based network with mature tools, then map its activations onto spike rates in an equivalent spiking network. Conversion is simple and reuses a huge toolchain, but tends to need many time steps (hence spikes, hence energy) to match the original accuracy, which partly undercuts the efficiency motive. Surrogate training generally yields sparser, lower-latency networks but is fussier to tune.
Learning on the implant itself
The most ambitious pillar is on-chip learning: adapting the decoder locally, in the implant, without shipping neural data to the cloud. Local rules — spike-timing-dependent plasticity, or three-factor / reward-modulated rules that gate plasticity by a global success signal — can be implemented in the synaptic devices themselves, so weights update where they are stored.
The honest caveat is that in-hardware learning stresses the very analog non-idealities of guide 3 even harder — asymmetric, nonlinear conductance updates make gradient-based rules misbehave, and demonstrations so far are small and mostly ex vivo.
The real chips, and where they run
Concrete neuromorphic platforms exist — digital spiking processors and mixed-signal chips among them, alongside research memristive prototypes. It is important to be precise about a fact the marketing often blurs: nearly all BCI demonstrations to date run the spiking decoder on such a chip outside the body, wired to recorded or streamed data, not on a chip inside the skull. The implanted neuromorphic decoder is still largely aspirational.
How do spiking decoders compare to the conventional decoders of Volume II? On simpler motor tasks, a well-trained SNN can approach the accuracy of an RNN or transformer-based decoder; on harder problems it typically still trails. The value proposition is therefore rarely raw accuracy — it is achieving comparable accuracy at a fraction of the energy and latency, which is exactly the currency the implant cares about.