How to determine overfitting, underfitting and best fit in deep learning? What about this graph?

Overfitting: train loss continues to decrease while test/val loss increases
Underfitting: train loss remains high and doesn’t decrease(not constant). can be better determined using accuracy on train set rather than loss.
Best fit: Best to choose the weights giving high accuracy(or any metric depending on the problem) on test/val set rather than finding it with loss.

In the above graph, you can choose the weights at Early Stopping Checkpoint point.

1 Like

how can i avoid underfitting and decrease the training loss?