shown for an "improved " version of Tanh(16) model which uses more convolutional filters per layer (32 instead of 25