Continual (lifelong) learning

Training a single model over an unending stream of changing data — new days, new tasks, new neural statistics — so that it keeps improving and accumulating capability rather than being retrained from scratch each time. For BCI the stream is the user's whole history with the device, and the goal is a decoder that integrates today's data on top of years of prior experience, exploiting what is stable while absorbing what has changed. This is the umbrella under which drift adaptation, catastrophic-forgetting mitigation, and long-term stabilization all sit.

The defining tension of continual learning is the stability-plasticity dilemma: a model plastic enough to track new statistics tends to overwrite old knowledge, while a model stable enough to retain it resists useful updates. The main tool families are regularization (penalize changing weights important to past performance), replay (interleave stored or generated past examples with new ones), and architectural growth (allocate new capacity for new conditions). Applying these on a physiological, non-stationary, safety-critical stream — with no clean task boundaries — is considerably harder than on the curated task sequences used in most machine-learning benchmarks.