idea:
Concretely, for each example in the the labeled training dataset xl, we forward propagate the example through a convolutional and a pooling layer to obtain the activation of the hidden units a(2). We now represent this example using a(2) (the “replacement” representation), and use this to as the new feature representation with which to train the softmax classifier.
design:
In this exercise, our goal is to distinguish between the digits from 0 to 4. We will use an “unlabeled” dataset with all 10 digits to learn the filters; we will then use a labeled dataset with the digits 0 to 4 with which to train the softmax classifier.
procedure: