Train Loss Versus Validation Loss
Training loss is measured on examples used for parameter updates. Validation loss is measured on examples held out from those updates.
Both matter, but they answer different questions:
training loss: can the model fit what it saw?
validation loss: does the fit appear to transfer?
If both losses are high, the model may be underfitting. If training loss is low but validation loss is high, the model may be overfitting.
Do not read a single point too strongly. Loss curves are patterns over time. A noisy validation curve can move up and down. The warning sign is a sustained gap or a sustained worsening of validation behavior while training behavior improves.
Enter 1 for training loss or 2 for validation loss: this loss is measured on examples not used for parameter updates.
Compute it first, then check your number.
Training loss is 0.1 and validation loss is 1.8. Enter 1 for no concern or 2 for possible overfitting.
Compute it first, then check your number.