Continual Learning Is Confusing

Continual learning is one of those topics I thought I understood until I actually tried to explain it out loud. The basic idea seems simple. A model keeps learning after it is deployed. It picks up new information over time instead of being frozen at the moment of training. Easy, right?

Not really. The more I read, the more I realize that most of the hard parts are not about learning new things. They are about not forgetting old things while you learn.

There is a phrase people use called catastrophic forgetting. The name is dramatic but it is accurate. When you train a model on new data, the weights that used to represent old knowledge get overwritten. The model does not know it is losing anything. It just quietly gets worse at things it used to be good at.

What is interesting to me is that humans do not seem to have this problem in the same way. If I learn a new language I do not forget how to ride a bike. Something about how we consolidate memory is very different from how a neural network updates its weights. I do not know enough neuroscience to say more than that, but it feels like an important clue that the current setup is missing something.

Most of the tricks I have seen so far feel like workarounds. Replay buffers where you mix old examples with new ones. Regularization that punishes the model for changing important weights too much. Freezing parts of the network. They all help but none of them feel like a real solution. They feel like patches on a fundamental design problem.

I keep wondering if the answer is going to involve separating memory from the model itself. Something more like an external store that the model reads from and writes to, rather than trying to cram everything into the same set of weights. That might just be my current bias though. It is easier for me to imagine because it looks like the kind of software I already know how to build.

The other thing I find confusing is evaluation. How do you even measure whether a continually learning system is getting better? You cannot use a fixed benchmark because the world it is learning from keeps changing. You end up needing to evaluate stability, plasticity, and forgetting all at the same time, and I have not seen a clean way to do that.

I do not have a strong opinion on where this is heading. I just know that when people say continual learning is almost solved, my instinct is that they are looking at a narrow slice of it. There is a lot of room here for someone to actually figure it out.