Figure 1. Setup for stabilizing an intracortical brain computer interface (iBCI) with adversarial domain adaptation.

(A) Initial iBCI decoder training on day-0. The decoder is computed to predict the motor outputs from neural signals, using either the full-dimensional neural recordings or the low-dimensional latent signals obtained through dimensionality reduction. This decoder will remain fixed over time after training. (B) A general framework for adversarial domain adaptation training on a subsequent day-k. The ‘Generator’ (G) is a feedforward neural network that takes day-k neural signals as the inputs and aims to transform them into a form similar to day-0 signals; we also refer to G as the ‘aligner’. The ‘Discriminator’ (D) is another feedforward neural network that takes both the outputs of G (aligned day-k neural signals) and day-0 neural signals as the inputs and aims to discriminate between them. (C) A trained aligner and the fixed day-0 decoder are used for iBCI decoding on day-k. The aligned signals generated by G are fed to the day-0 decoder to produce the predicted motor outputs.