assertion failed vl_simplenn.m (see also #34) about matconvnet

[lenck]

Hi, sorry for such a late response.
The issue #34 is mainly about the difference between softmax and softmaxloss. softmaxloss layer is in fact sort of two layers - a softmax with a logistic loss on top.

So, in case of the issue #34, when you want to output the class probabilities, softmax is the way to go, whereas when you train, you usually know the GT labels and you want to optimise some sort of loss based on the mistakes the network makes. So, then, softmaxloss is the friend here as it does both :)

But of course, you can also use the softmax on top even for backpropagation, but then you would have to add the loss layer on top of that... Honestly, I get lost in this quite often too (I just hope that I'm correct ;)), but for example I quite like e.g. this explanation of softmax.