Representational drift

The slow, ongoing change in how individual neurons respond to the same stimulus, action, or task over days to weeks, even when behavior and performance are stable. A neuron that fires for a rightward reach today may fire less, differently, or not at all next week, while the population as a whole continues to represent rightward reaches. Drift has been documented most strongly in hippocampus, posterior parietal cortex, olfactory (piriform) cortex, and to a more debated degree in sensory and motor cortex; its causes (ongoing synaptic turnover, plasticity, changing internal state) are still not settled.

For BCI this is the central adversary: a decoder is a fixed map from measured neural features to intended output, and drift silently invalidates that map. Crucially, the observed day-to-day change in a real recording is a mixture of genuine biological drift and non-biological recording instability, and the two are hard to separate at the electrode. A key empirical finding softens the picture — although single-neuron tuning drifts, the low-dimensional population dynamics (the neural manifold) are often far more stable, which is what makes latent-space stabilization and drift-robust decoding feasible at all.