I am running a neural net on a tabular dataset. The output is from 10 epochs of a tabular_learner model using fit_one_cycle.
My question is, with every additional epoch I add is increasing the overall accuracy. However, the train_loss is lowering to an extremely low number. Is this an indication of overfitting? I remember Jeremy saying that as long as the accuracy is continuing to improve that it’s very unlikely that you are overfitting, but I can’t understand how the train loss can be so low.
Yes it is overfitting. It’s learning how to best do with the training set and not the validation set. A good experiment would be to have another test set and explore how the model performs on the test set when the train and valid loss are the same, or when one is much larger than the other like you see.
I see that the valid_loss is worsening - but how come the accuracy is improving? Surely the accuracy will get worse if its overfitting to the training set and therefore predicting wrong on the validation_set?