ABSTRACT

In this chapter, the authors perform the same character transformations by taking advantage of the fact that they have the whole spectrum of real numbers at our disposal. In general, it is not possible to preserve one-hot encodings on the output by doing the kind of compression they did, using only one decoding layer to decompress it. The compression here destroys the independence of the input dimensions, and it is not possible to recover that independence with just one decoding layer. The term deep learning is associated to neural networks that have more than one hidden layer. In contrast, both the initial bow tie network and the neural networks are called shallow networks.