Well, on a high level, an autoencoder can be thought of as a specific type of feed-forward network that learns to mimic its input to reconstruct a similar output. As we mentioned previously, it is composed of two separate parts: an encoder function and a decoder function. We can think of the entire autoencoder as layers of interconnected neurons, which propagate data by first encoding its input and then reconstructing the output using the generated code:
The previous diagram illustrates a specific type of autoencoder network. Conceptually, the input layer of an autoencoder connects to a layer of neurons to funnel the data into a latent space, known as the encoder function. This function can be generically defined as h = f(x), where x refers to the network inputs and h refers to the latent space that's...
