FREE · NO ACCOUNT REQUIRED
Prediction: learn & practise
Interactive learning · Neural networks
Watch a network learn.
Change the architecture, train on synthetic data and inspect how its predictions evolve. Open the research workspace
From examples to a decision boundary
A neural network transforms inputs through weighted connections. Training updates those weights to reduce binary cross-entropy loss. Choose tanh, ReLU, sigmoid or linear hidden activations; the output uses a sigmoid probability.
- Choose a pattern and architecture.
- Start training or advance one epoch at a time.
- Compare training loss with loss on a separate validation sample.
The coloured background is the probability of the inner/positive class. Dots show held-out validation examples. This small network uses full-batch gradient descent; each epoch updates weights once using the selected training subset.
Learning progress is saved on this browser when storage is available. Return to Academy.
Watch the decision boundary bend around the inner class.
Changing data or architecture resets training. Training stops at 1,500 epochs.
Interpret the experiment
Lower training loss alone does not prove better generalisation. Repeated tuning against the same validation sample makes it part of model selection. Real research needs an independent final holdout (available here as a one-time evaluation per configuration), representative data and appropriate uncertainty analysis.
Final holdout is hidden until you freeze this run. Evaluate once after model selection.
Export includes settings, scores and the complete loss history. Synthetic data and fixed seeds make this a reproducible teaching exercise, not evidence of performance on your research data.
Method reference: Goodfellow, Bengio & Courville — Deep Feedforward Networks.