Improving neural network representations using human similarity judgments