Autoencoding is a data compression and decompression algorithm implemented with an ANN. Since it is an unsupervised form of a learning algorithm, we know that only unlabeled data is required. The way it works is we generate a compressed version of the input by forcing it through a bottleneck, that is, a layer or layers that are less wide than the original input. To reconstruct the input, that is, decompress, we reverse the process. We use backpropagation to both create the representation of the input in the intermediate layer(s), and recreate the input as the output from the representation.
Autoencoding is lossy, that is, the decompressed output will be degraded in comparison to the original input. This is a similar situation to the MP3 and JPEG compression formats.
Autoencoding is data-specific, that is, only data that is similar to that which they have been trained...
