What are autoencoders?

Autoencoders are a type of deep NN (DNN) that can learn an efficient reduced representation of the data in an unsupervised way and minimize the error between the compressed and subsequently reconstructed data compared to the original data. Why compress the data and then reconstruct the original data? Isn’t it counterintuitive? Suppose you are on your vacation and took pictures, but you realized after you return from vacation that a picture has noise because of dim light. Wouldn't it be nice if there was a way to remove the background and make the picture great? This is, in computer vision lingo, called feature variation, which removes any noise in a picture. This is what autoencoders do. They learn a representation or latent space from the training data to ignore signal noise. The compression step forces the network to only learn the most important latent features. This is because if the model is at full capacity, it will just copy the data without...