An implanted decoder learns to read neural activity, and then the activity changes. Electrodes shift microscopically, tissue responds, and the mapping the decoder learned yesterday is slightly wrong today. Left alone, performance drifts downward.

Unsupervised adaptation is the obvious answer, since nobody wants to recalibrate a brain implant every morning. The catch is that existing methods use deep recurrent or adversarial architectures, and an implantable device has a power budget measured against tissue heating. Adaptation that requires a workstation cannot run where it is needed.

Membrane Potential Alignment realigns a pretrained spiking decoder to the shifted recordings, and the choice of what to align is the elegant part: membrane potential is already the internal state the network maintains, so the correction operates on a quantity the hardware computes anyway rather than requiring a second model on top.