Interactive machine learning · 2026
Learning Rule Lab
Small experiments that make learning visible: follow the weights, compare update rules, and see what a network forgets.
Two ways to learn from the same examples
Hebbian updates
Fruit features and the expected class strengthen their connections together. The visualization exposes the individual weights rather than hiding the update inside a library.
Backpropagation
Prediction error flows back through the network to adjust the weights. Training-step responses let the browser show changes in predictions, loss and connections.
A neuron finds a direction
An Oja neuron learns the first principal component of a two-dimensional dataset. The browser displays the learned direction alongside the data, connecting a weight update to a geometric interpretation.
- Input
- A small two-dimensional dataset with a dominant direction.
- Update
- Observe how the neuron’s weights move as it learns.
- Geometry
- Compare the learned vector with the data’s principal axis.
What happens when the task changes?
- Learn Task ATrain a small ReLU network on the first classification task.
- Switch to Task BContinue learning and observe how performance on Task A changes.
- Compare EWCUse an empirical Fisher estimate to penalize changes to parameters important for the earlier task.
Built to explain the mechanism
The experiments use small datasets and inspectable calculations. They illustrate how learning rules behave; they are not a benchmark of large-scale continual learning.