The deep autoencoder
	784  1000   500   250                                                          30 linear units      784  1000   500   250     If you start with small random weights it will not learn.  If you break symmetry randomly by using bigger weights, it will not find a good solution.    So we train a stack of 4 RBM’s and then “unroll” them.  Then we fine-tune with gentle backprop.