OpenAI's Will Depue proposes using a localized distance penalty during training to improve the visual interpretability of weight matrices · Digg