Continual Backprop

Continual Backprop is a training technique that allows an AI model to keep learning new things even after it has finished its initial training, while making sure it doesn't forget what it learned before.

Here's an analogy with the human brain. As you learn new things, some brain cells can end up so "used up" that they stop picking up new patterns well. Continual Backprop works in a similar way: it gradually identifies parts of the network that have become stiff and less useful, resets them back to an initial state, and lets them learn again. The rest of the network is left untouched, so previously learned knowledge doesn't get wiped out all at once.

Most AI models today stop changing once training is done — even as they gain new experience out in the world, they don't update themselves. Richard Sutton, widely regarded as a founder of reinforcement learning, argues that without genuine continual learning, an agent can never truly learn about and correct its understanding of the world on its own. Continual Backprop is one attempt to enable that kind of continual learning while avoiding "catastrophic forgetting" — the loss of everything learned before.

In articles, this is introduced as a technique published in Nature by Richard Sutton's research team. Note that it's not the name of a specific product or service, but a research term referring to a method for making models keep learning even after training has ended.