Training the auto encoder model is a bit tricky, hence a detailed illustration has been provided for better understanding. During the training phase, the whole encoder-decoder section is trained against the same input as an output of decoder. In order to achieve the desired output, features will be compressed during the middle layer, as we are passing through the convergent and divergent layers. Once enough training has been done by reducing the error values over the number of iterations, we will use the trained encoder section to create the latent features for next stage of modeling, or for visualization, and so on.
In the following diagram, a sample has been shown. The input and output layers have five neurons, whereas the number of neurons has been gradually decreased in the middle sections. The compressed layer has only two neurons, which is the number of latent dimensions we would like to extract from the data.
The following...