negative-self-influence. neural networks don't minimize loss [caution: probably due to batchnorm]