2 ms·
> using max(0, exp(x)-1) instead of exp(x) Won't this cause the gradient to vanish on the left half, causing problems with training?
by smallnamespace 2y ago
> using max(0, exp(x)-1) instead of exp(x)
Won't this cause the gradient to vanish on the left half, causing problems with training?
- espadrine 2y agoThat is a concern that is shared with ReLU. But since the weights are shared across the context/minibatch, perhaps that would not be an issue, similar to ReLU.