Hands-On Neural Networks with Keras

What's the catch?

A problem with all of the networks we have built so far is that they only accepted inputs and outputs of fixed sizes for given training examples. We have always had to specify our input shape, defining the dimensions of the tensor entering our network, which in turn returns a fixed size output in terms of a class probability score, for example. Moreover, the hidden layers in our networks each had their own weights and activations, which behaved somewhat independently of each other, without identifying relationships between successive input values. This holds true for both the feedforward and the CNNs that we have familiarized ourselves with in previous chapters. For each network we built, we used non-sequential training ...