3 ms·
> Actually, there's compelling evidence that overfitting is a necessary step for achieving state-of-the-art performance with DNNs! I think the term you are loo
by trott 6y ago
> Actually, there's compelling evidence that overfitting is a necessary step for achieving state-of-the-art performance with DNNs!
I think the term you are looking for is "overparametrization". Overfitting is, by definition, worse generalization performance despite better training performance.