You can get (lossy) compression by using an autoencoder with a hidden layer that has fewer nodes than the number of inputs, since then you only need to store th

You can get (lossy) compression by using an autoencoder with a hidden layer that has fewer nodes than the number of inputs, since then you only need to store the hidden layer excitations for each example. The weights and network architecture itself are part of the compressor, not the compressed data. As for how well this works compared to other approaches, I don't know, and this blog post certainly doesn't address the question.