Learning Paradigms

continual learning

/ kun-TIN-yoo-ul LER-ning /

People learn new things all their lives without wiping out the old: learn to drive a van, and you don't forget how to drive a car. Machines are surprisingly bad at this. Continual learning — also called lifelong learning — is the effort to build models that can keep learning new tasks over time while holding on to what they already knew.

The obstacle has a dramatic name: catastrophic forgetting. A neural network stores everything in one shared set of weights. Train it on a new task and the weights shift to fit the new data — overwriting the very settings that encoded the old task, which can then collapse almost completely. So continual learning is mostly about defending old knowledge while absorbing new: replaying a few stored old examples, protecting the weights that matter most to past tasks, or setting aside separate capacity for new tasks. The ideal is to learn from a stream of changing tasks and end up competent at all of them.

This matters because the real world doesn't sit still — new product categories, new slang, new fraud patterns keep arriving, and retraining a giant model from scratch every time is wasteful or impossible. The honest state of things: continual learning is an active, unsolved research area. Every method trades something off — memory to store old data, compute, or a ceiling on how much new stuff fits — and none yet matches the easy, graceful way humans accumulate skills over a lifetime.

A model trained to recognize ten animal species is later trained on ten new ones — without revisiting the originals. By default it now nails the new ten but flubs the original ten badly: catastrophic forgetting. Mixing in a small replay buffer of old examples during the new training keeps the original skills mostly intact.

Learning the new can erase the old — that's catastrophic forgetting.

The core enemy is catastrophic forgetting: training on new tasks overwrites the shared weights that held old knowledge. No method fully solves it yet — each one trades off memory, compute, or how much new learning fits.

Also called
lifelong learningincremental learning持续学习持續學習终身学习