Training improves, validation worsens

Foundation, about 6 minutes. Question R012.

During training, training loss keeps falling while validation loss begins rising. What is a plausible explanation?

Answer it in the app It's free. You get hints, a second try if you miss, and a follow-up question.

Show the answer and worked solution

Answer: D. Training-specific patterns are improving fit but hurting generalization.

Training and validation losses measure performance on different observations.
Fitting increasingly sample-specific patterns can lower training loss while raising validation loss.
This is consistent with overfitting, not proof of it; noisy metrics and distribution differences still need checking.

Interview shortcut. Training loss falling while validation loss rises is the overfitting signature.

Takeaway. Use validation behavior to examine generalization, without treating a loss curve as a complete diagnosis.

Why the other choices fail

The diverging losses show that gradient updates have ceased. Continued training-loss changes are inconsistent with explaining the pattern solely by stopped parameter updates.
The learning rate is too low, and the optimizer has stalled on validation. A too-small model would underfit both losses.
A smaller training error necessarily implies a smaller population error. Training error measures the objective on the fitted sample. It can improve while held-out performance worsens.

In the app this has a follow-up where one thing about the setup changes. Try the follow-up.