Randomness in AI Training Pushes Autoencoders to Favor Large Decoder Weights

Original: Broken scale symmetries in undercomplete linear autoencoders

Source: arxiv.org ↗