Funnily enough, some over-fitting is nearly always a good thing. All that matters in the end is: is the validation loss as low as you can get it (and/or the val accuracy as high)? This often occurs when the training loss is quite a bit lower.
41 Likes