Validation loss

Validation loss is the error a model scores on a held-out validation set — data it isn’t trained on — measured with the same loss function used during training. It’s the counterpart to training loss (error on the data the model is actively learning from). The gap between the two is the headline diagnostic when training a machine learning model: while both fall together, the model is genuinely learning; when training loss keeps dropping but validation loss flattens and then starts rising, the model has begun overfitting — memorizing the training data instead of generalizing, often a sign it has more parameters than the data can constrain. That turning point is the usual signal for early stopping (keeping the checkpoint at the lowest val loss) and for tuning regularization, learning rate, or dataset size. Crucially, the validation set is for monitoring and tuning only; a separate test set is reserved for the final, unbiased performance estimate.