Recording instability

The non-biological sources of day-to-day change in a chronic recording: micromotion of the array relative to tissue, growth of a glial scar that raises impedance and attenuates signals, gradual electrode corrosion or encapsulation, single-unit dropout as isolable neurons are lost and new ones appear, spike-sorting instability, and reference or amplifier changes. These produce shifts in the measured feature distribution that look, to a decoder, much like biological drift — but their remedy is different (better spike sorting, threshold-crossing rather than sorted units, impedance-robust features, or re-anchoring to stable channels).

Separating recording instability from representational drift matters because it changes what a fix can hope to achieve. If a channel is simply gone, no amount of latent alignment recreates its information; if the population code has merely rotated, alignment can recover it. Many drift-robust methods deliberately sidestep the distinction by working on population-level or threshold-crossing features that are more forgiving of unit turnover than carefully sorted single units.