Overfitting — read the learning curves

Background

The signature of overfitting is a diverging pair of curves: the training loss keeps falling while the validation loss turns back up. The model is still improving on examples it has seen, but getting worse on held-out data — it's memorising noise.

Problem statement

Implement is_overfitting(train_losses, val_losses) that returns whether the most recent epoch shows the overfitting signature: training loss decreased but validation loss increased from the previous epoch.

Input

• train_losses — list of per-epoch training losses (length $\ge 2$).
• val_losses — list of per-epoch validation losses, same length.

Output

Returns a bool: True if train_losses[-1] < train_losses[-2] and val_losses[-1] > val_losses[-2], else False.

Examples

Example 1 — classic overfitting

Input:  train_losses = [0.5, 0.3], val_losses = [0.4, 0.5]
Output: True

Explanation: training improved ($0.5\to0.3$) while validation worsened ($0.4\to0.5$).

Example 2 — both still improving

Input:  train_losses = [0.5, 0.3], val_losses = [0.5, 0.4]
Output: False

Explanation: validation is still dropping, so no overfitting signal yet.

Constraints

• Compare only the last two epochs.
• Both conditions must hold (train down and val up).
• Return a Python bool.

Notes

• This local check is the trigger behind early stopping: when it fires persistently (over a patience window), training has started baking in noise.