Model Building & Algorithm Learning

Neural Network Playground

Build a small multilayer perceptron, train it one epoch or many, and watch the loss, activations and decision boundary evolve.

7 · Regression 8 · Clustering 9 · Decision boundaries 10 · Neural networks 11 · Gradient descent 12 · Ensembles 13 · Step-by-step

Lab concept guide

What to observe while you experiment

A multilayer perceptron composes affine transformations and nonlinear activations. Training repeatedly computes predictions, loss and gradients, then updates weights; width/depth affect representation capacity while optimisation and regularisation affect what is actually learned.

MechanismStep through epochs and connect loss changes, hidden activations and decision-boundary changes to weight updates.

Failure modeInterpreting training loss decrease as guaranteed generalisation or adding layers/units without checking data size, validation behaviour and optimisation stability.

VerificationTrack both training and validation loss/metrics, inspect a few activations/predictions, and verify that one epoch actually changes weights and outputs.

Experiment deliberately

Train one epoch at a time. Predict how increasing hidden units or learning rate should affect capacity/optimisation, then compare the loss and boundary.

DatasetHidden layers2Neurons / layer5ActivationLearning rate0.060Seed

Network initialised.

Decision surface

Learning curve

Network & probe activations

Probe xProbe y