Neural Network Capacity Explorer

Train a small tabular MLP and watch training AUC climb with capacity while internal test AUC refuses to follow.

  • neural network
  • overfitting
  • capacity
  • generalization gap
Training AUC
0.833
Internal test AUC
0.802
External test AUC
0.744

Training loss

05010015020025030035040000.10.20.30.40.50.60.7EpochCross-entropy

Generalization gap

chance00.511.522.5300.20.40.60.81AUC

Capacity sweep: hidden units vs. AUC

5101520253000.20.40.60.81Hidden unitsAUC
TrainingInternal test

Model comparison

DatasetAUC
Training0.833
Internal test0.802
External test0.744

AUC on internal and external test sets.

Capacity is not the bottleneck
With roughly a hundred training cases and eight tabular predictors, increasing hidden units drives training AUC toward 1.0 while internal test AUC stays flat or drops. This gap — the generalization gap — is the single most important diagnostic in deep learning. No combination of learning rate, epochs, or regularization manufactures signal that the data do not contain; deep networks earn their keep on raw voxels with thousands of studies, not on a handful of tabular features.