I am doing a multiclass classification using neural networks. say I have 10 target classes and one null (non-of-the-above-targets). is it better that I train a neural network separately for each target with two output neurons for each network (target , non-target) so that i would need 10 separate neural networks in the case above or a neural network with 11 outputs (10 targets, 1 -other) ? i mean i have seen people using either of the two approaches in different papers but without explanation. is there a theoretical superiority to using a separate network for each target class ?is the computational overhead cost worth the gain and benefit with respect to the alternative approach ?
Thanks in advance!
P.s. 1: of course in either of the approaches the distribution of training examples is heavily skewed towards the non-target ("other") class . 2: the output layer of the NN is assumed to have a softmax activation.