Fast or slow training? A training-curve game
Every chart in this game is a real training run of Lecture 2. You see how the error of the network evolves, epoch by epoch, and you say which setting produced the curve: the learning rate, the batch size or the optimiser. Nine charts, about three minutes.
The first six runs fit a multilayer perceptron (MLP 64-64) to the 20,640 Californian census block groups of the lecture; the chart shows the mean squared error (MSE) on 4,128 validation block groups the network does not train on (lower is better). The last three runs detect card fraud; the chart shows the area under the precision-recall curve (PR-AUC) on the validation transactions (higher is better). There is no model to beat here: only your score counts.
What the game shows
- With plain stochastic gradient descent (SGD) and 30 epochs, the learning rate decides everything: η = 0.0001 leaves the validation MSE at 1.093, η = 0.1 reaches 0.281, and from η = 0.3 upwards training diverges (median of 3 seeds).
- The optimiser sets the speed: after one epoch on card fraud the validation PR-AUC is 0.004 with plain SGD, 0.488 with momentum and 0.592 with Adam; on the test transactions the three end at 0.786, 0.823 and 0.803 (5 seeds).
- Small batches learn more per epoch (validation MSE after 5 epochs: 0.294 with batches of 16, 0.870 with 1,024) but cost more time: 25.8 s against 1.0 s for 40 epochs.
Data: California housing, 1990 census (Pace and Barry, 1997; StatLib), split 60/20/20 at random; card fraud of the Université libre de Bruxelles (ULB) and Worldline, Open Machine Learning platform (OpenML) dataset 1597, split by time. Curves: the runs of analysis_L2.py (seed 20261108; cached in lectures/L2/cache). Code: lectures/L2/make_game_L2.py.

