Learning Curves

A learning curve shows how a metric changes during training.

Curves help answer questions a final number hides:

  • did training ever improve?
  • did validation improve with training?
  • did loss explode?
  • did progress stop early?
  • did validation worsen while training improved?

The shape matters. A single final metric cannot tell whether the run was healthy, unstable, or lucky.

Learning curves are especially useful when comparing changes. If a new learning rate reaches lower validation loss faster, the curve shows that. If a new model has the same final score but unstable training, the curve shows that too.

Exercise: Stuck curve

Validation loss stays near 1.0 for many epochs while training loss also stays near 1.0. Enter 1 for likely progress or 2 for likely stuck.

Compute it first, then check your number.

Exercise: Useful curve evidence

Enter 1 if only the final metric matters, or 2 if the path of training can reveal instability.

Compute it first, then check your number.