Features
Change the inputs before changing the network.
Start with a simple dataset and toggle x1, x2, squared terms, products, and sine features one at a time. Watch which feature makes the decision boundary suddenly easier to draw.
Neural Network Playground
A self-hosted build of TensorFlow Playground. Adjust datasets, features, hidden layers, activations, and regularization to see how a small neural network learns.
Lab guide
Use the embedded playground below as a controlled experiment: change one concept, run training, then explain the shift in the boundary before touching the next control.
Hidden layers
Add capacity slowly.
Increase width or depth by one step, then train again. Look for the moment the model stops underfitting, and the moment extra neurons only make the boundary wigglier.
Local minima
Run the same setup more than once.
Reset and retrain without changing controls. If the loss plateaus differently, compare the paths and then test whether learning rate, activation, or regularization helps the model escape.