Five rounds on why a network that learns its training data perfectly can still fail, and what stops it. In each round you first predict what will happen and lock your answer, then test it on the course data, and only then see the proof. The data are those of the lecture: 1,000 German credit applicants, the card fraud sample (time-ordered split) and California housing.
Charts stay empty until you lock a prediction. In round 3, drag the bars to turn them; scroll or pinch to zoom.
Round 1 · Memorising
- 1 Predict
- 2 Test
- 3 Prove
Round 2 · Loss or ranking?
- 1 Predict
- 2 Test
- 3 Prove
Round 3 · Dropout or weight decay?
- 1 Predict
- 2 Test
- 3 Prove
Round 4 · When to stop
- 1 Predict
- 2 Test
- 3 Prove
Round 5 · Beyond the interpolation point
- 1 Predict
- 2 Test
- 3 Prove
German credit: 600 training, 200 validation and 200 test applicants, 48 standardised inputs; network 256-256 with ReLU, AdamW (learning rate 0.001), batch 64, 150 epochs, no early stopping; logistic regression with C = 1. Card fraud: 28,924 transactions in time order (60/20/20), network 256-256, batch 256, 120 epochs; average precision with the weights that restore the full-data class mix. Double descent: 400 training and 5,000 test block groups per draw, random ReLU features, 10 draws, median. Seeds 20261108 + s. Code: lectures/L3/analysis_L3.py and interactive_data_L3.py on GitHub.

