Catastrophic forgetting

The tendency of a neural network trained on new data to abruptly lose competence on earlier data, because gradient updates for the new distribution overwrite the weights that encoded the old one. In a drift-robust BCI this is the failure that a naive fix invites: retrain the decoder each day only on today's data and it may decode today well but fail on the range of conditions it once handled, so accuracy becomes brittle and swings with whatever was seen most recently.

The standard mitigations map onto the continual-learning toolkit. Regularization methods such as elastic weight consolidation add a penalty that anchors weights deemed important to past performance (estimated from the Fisher information), letting less-important weights move freely. Replay or rehearsal interleaves a buffer of past examples with new ones so old distributions are never absent from training. Parameter isolation freezes or grows dedicated capacity per condition. In BCI these must run under tight compute, latency, and privacy limits, and the absence of labeled past data can make replay itself hard to implement faithfully.