Picture a robot in a warehouse. For a year it has picked red boxes off a belt, and it is good at it. On Monday somebody teaches it the new blue boxes. On Tuesday it has become an expert in blue… and looks at the red ones as if it had never seen them.
Nothing is broken. This is how most neural networks behave when you teach them one thing after another. The problem has a name, catastrophic forgetting, and I have spent a good part of the last 25 years on it.
The short answer
Continual learning is the ability to learn new things over time, one after the other, without retraining from scratch and without losing what was learned before. Animals do it from birth. Most AI does not, and for a machine that works in the physical world that is a daily problem.
Why networks forget
A deep network stores what it knows in millions of connections, and every piece of knowledge is spread over many of them. When you train on something new, the same connections get adjusted. Nothing tells the network which ones were holding the old knowledge, so it writes over them.
In 1989 two psychologists, McCloskey and Cohen, showed it with arithmetic: teach a network one set of additions, then a second set, and the first is gone. They called it catastrophic interference. The name stuck because it fits.
The standard fix is to retrain on everything, old and new together. It works. It also means keeping all the old data, a lot of computing, and a trip back to the data center each time the world changes.
Why it matters more for robots
A model that answers questions can be retrained every few months and nobody notices. A robot, a drone or an inspection camera lives in a place that changes every day: a new part on the line, different light in winter, a pallet where there was none yesterday.
- Things change faster than retraining. By the time the new model ships, the line has changed again.
- The data is local. Often it cannot leave the factory, or there is no connection to send it over.
- The device is small. Learning has to fit in the power and memory of the machine that does the job.
This is why I think of continual learning as part of Physical AI, and not as a footnote of machine learning.
What biology does
You learned a new face this morning and you still remember your mother’s. Brains had to solve this early… an animal that forgets what a predator looks like each time it learns a new food does not leave many descendants.
Two ideas from neuroscience are worth knowing. The first is that the brain has a fast learner and a slow one. The hippocampus picks up new events quickly, then replays them to the cortex over time, so the new is woven into the old instead of written over it (McClelland, McNaughton and O’Reilly, 1995).
The second is Adaptive Resonance Theory, from Stephen Grossberg and Gail Carpenter (Carpenter and Grossberg, 1987). Here learning happens only when what comes in matches what the system expects. If the match is poor, the system opens a new category and leaves the old ones alone. I trained in that school, and it shaped how I build things.
Three ways engineers fight forgetting
- Replay. Keep some old examples, or generate them, and mix them with the new ones. Close to what the hippocampus does.
- Protection. Work out which connections mattered for the old task and make them harder to change. The best-known version is elastic weight consolidation (Kirkpatrick et al., 2017).
- Structure. Give new knowledge its own room: new units, new categories, or a part of the network that learns fast while the rest stays put.
None of them is free. Replay needs memory, protection gets harder as tasks pile up, and structure grows. In practice you mix them, and you pick the mix that fits the device.
What I did about it
At Neurala, the company I co-founded, we built this into a product. The approach, which we called Lifelong-DNN, keeps a deep network for the heavy lifting of seeing, and adds a fast-learning part on top that can take in a new object from a few examples, on the device, without going back to the data center. It ended up in drones, in cameras and on factory lines.
I wrote about the details over the years:
- What happens when AI never stops learning?
- Overcoming the Catastrophe of Forgetful AI
- Learning That Never Stops
Questions I get
What is continual learning?
Continual learning is the ability of a system to learn new things over time, one after the other, without retraining from scratch and without losing what it learned before. It is also called lifelong learning.
What is catastrophic forgetting?
Catastrophic forgetting is what happens when a neural network is trained on something new and, in the process, overwrites what it already knew. The term goes back to a 1989 paper by McCloskey and Cohen, who called it catastrophic interference.
Why does continual learning matter for robots?
A robot works in a world that keeps changing: new objects, new lighting, new tasks. If every change means sending data away and retraining the whole network, the robot is always a step behind. A robot that learns on the spot, and keeps what it knew, is useful for much longer.
How do you reduce catastrophic forgetting?
There are three main families of methods. Replay: mix old examples, or generated ones, with the new. Protection: make the connections that mattered for old tasks harder to change. Structure: give new knowledge its own room in the network, and learn only when the input matches what the system expects, as in Adaptive Resonance Theory.