The interaction between the encoder and decoder functions is governed by yet another function, which operationalizes the distance between the inputs and outputs of the encoder. We have come to know this as the loss function in neural network parlance. Hence, to train an autoencoder, we simply differentiate our encoder and decoder functions with respect to the loss function (typically using mean squared error) and use the gradients to backpropagate the model's errors and update the layer weights of the entire network.
Consequently, the learning mechanism of an autoencoder can be denoted as minimizing a loss function, and is as follows:
min L(x, g ( f ( x ) ) )
In the previous equation, L represents a loss function (such as MSE) that penalizes the output of the decoder function (g(f( x ))) for being divergent from the network's input, (x). By iteratively...
