Day 40: Train/val curves: diagnosing over- and under-fitting
Reading the curves like an X-ray
The train/val loss and accuracy curves are your diagnostic instrument. The *shape* of the gap between them tells you exactly what's wrong and what to do — no guessing. This is the skill that turns training from trial-and-error into deliberate iteration, and it's an explicit Stage 1 exit criterion ('train/val curves shown').
A healthy run: both rise together and the gap stays small. A widening gap would signal overfitting.
The three diagnoses
- Both low, still rising → underfitting or undertrained: train longer, unfreeze more layers, or use a bigger model.
- Train high, val plateaus then gap widens → overfitting: add augmentation/dropout/weight decay, or get more data.
- Both high, gap small → healthy: you're done, or push a little further carefully.
The curve also catches boring bugs
A loss that's flat from step one usually means a broken learning rate or a data/label mismatch, not a hard problem. A validation curve *better* than training often means you forgot model.eval() (dropout still active in training inflates train loss). The curves diagnose bugs as readily as they diagnose fit.
Key terms
- Learning curve
- A plot of a metric (loss or accuracy) over training epochs for train and validation sets.
- Generalization gap
- The difference between training and validation performance; its size and trend diagnose over/underfitting.
Training accuracy climbs to 98% while validation accuracy plateaus at 78% and the gap keeps widening. What is the diagnosis and a correct fix?
