(no subject)
I so had this today. Correlations of correlations actually can tell something useful about the hidden layer. (Too bad that, in this case, the hidden layer's contribution to the converging input of the other layer appears non-selective, but hey, you can't have it all...)
