Learning from compressed observations