PYTHON · NUMPY

Catastrophic forgetting

Watch the failure that keeps today's models frozen: a network learns a second task by overwriting the first.

01Challenge

Train one network on two tasks in sequence and measure how much of the first it destroys learning the second.

Predict firstTrain a standard neural net on task A, then train it on task B. What happens to task A?

02Model

An embodied agent meets its skills in sequence — a new tool this week, a new surface next week. That is exactly where standard networks break. Train a network on task A until it is good, then train the same network on task B, and its accuracy on A collapses: the weights that encoded A get overwritten to fit B. This is catastrophic forgetting, and it is a large part of why frontier models are trained once, on everything at once, and then frozen — they cannot safely keep learning on the job. Here you measure it directly. You will teach one network task A, confirm it learned, teach it task B, and watch task A fall apart. Naming the failure precisely is the setup for the fix: the interference comes from every task sharing the same weights. The next lesson stops them from sharing.

03Guided practice

1. Two tasks. Each a 2D split along a different boundary — skills that arrive one after another. 2. Learn in sequence. Train on A, measure A; train on B, measure B and A again. 3. The damage. Task B is learned, but task A has collapsed toward chance. One shared network, and the new task wrote over the old.

04Feedback

PASS WHEN PASS when both tasks are learned in turn and task A's accuracy collapses after learning B.

If your run fails, check:

A cell errored — run them in order.
Task A should drop sharply after the network learns task B.

05Retrieve & space

What is catastrophic forgetting? (learning a new task overwrites the weights that held an old one).
Why does it happen in a standard network? (every task shares the same weights).
What real limitation does it impose? (models are trained once and frozen, not updated on the job).

06Mastery & project

You can reproduce and measure catastrophic forgetting, and name its cause as shared weights.

Before asking a deployed agent to learn something new, ask what it will forget to make room.

← The neuron is a network Dendrites that don't forget →

PYTHON · NUMPY

Catastrophic forgetting

Learn task A, then task B, and measure the forgetting.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you