MIMO nonlinear memory model

The computational core of the hippocampal memory prosthesis: a multi-input, multi-output (MIMO) nonlinear dynamical model that predicts the spatiotemporal spike pattern one hippocampal subfield (e.g. CA1) will emit given the pattern arriving at another (e.g. CA3). Because the input-output transformation is nonlinear, non-stationary, and involves many interacting channels, it is typically formulated as a sparse generalized Laguerre-Volterra model — a Volterra-series expansion whose kernels are compactly represented in a Laguerre basis and estimated under sparsity constraints — or, more recently, with related point-process and machine-learning regressions.

Crucially the model is estimated per patient and can be inverted to compute the stimulation pattern needed to reproduce a desired output, which is what turns a descriptive model into a prosthetic controller. Its limitations are those of any data-driven identification: it is only as good as the recorded training data, assumes the recorded channels capture the functionally relevant signal, and does not model plasticity or slow drift.