Zero-bias autoencoders and the benefits of co-adapting features