The neural code problem

A BCI can only be as good as our understanding of what neural activity means. Yet the neural code, how populations represent variables, over what timescales, in rate versus timing versus population geometry, is only partially understood and differs across areas, tasks, and individuals. Decoders often sidestep this by learning a statistical mapping without a mechanistic code, which works for well-practiced motor outputs but stalls for abstract cognition, where we cannot label the target variable, let alone read it.

The frontier view is that latent-dynamics and foundation-model methods may let us decode useful structure before we understand it, but that writing and higher cognition will remain gated by genuine gaps in coding theory. Whether a universal, low-dimensional description of neural computation exists, or whether the code is irreducibly high-dimensional and idiosyncratic, is an open scientific question with direct engineering stakes.