> + Reconstruction in deep NNs is often guided by an l2 distance. If there are competing feature sets, gradient descent moves the configuration to one of the st

> + Reconstruction in deep NNs is often guided by an l2 distance. If there are competing feature sets, gradient descent moves the configuration to one of the stabilizers. > - Note. Is this indeed the case? Is there a WTA mechanism at play? Is this preferable? > - Note. In this view a NN is a MCMC that is stopped prematurely. Generally the 'performance' of the autoencoder is measured at the reconstruction step, i.e. what is the reconstruction accuracy of the autoencoder? This is an L2 loss function.