Online Learning

Update the model as data arrives, rather than retraining on a corpus every so often. The data comes once, in the order it happened, and cannot be shuffled or revisited — which rules out most of what ordinary training assumes and is exactly the situation a recommender or a fraud model is in.

The distinction from Continual Learning is what is changing. That is about a sequence of tasks and the risk of forgetting the old ones; this is one task with a moving distribution, where the risk is instability — a burst of unusual traffic moves the model, and by the time anyone notices, the thing that produced the burst is over. Almost everything deployed is a hybrid: frequent small updates on a base that is retrained from scratch on a slower cycle, so drift has a floor to return to.

The learning rate is doing two jobs at once and they conflict. High enough to track a genuine change is high enough to be moved by noise, and low enough to ignore noise is low enough to lag a real shift — with no setting that does both, because the two are the same signal at different timescales. Everything here is a way of estimating which one you are seeing, and none of them can do it from the current sample alone.