Local Explanation Methods for Deep Neural Networks Lack Sensitivity to Parameter Values

Explaining the output of a complicated machine learning model like a deep\nneural network (DNN) is a central challenge in machine learning. Several\nproposed local explanation methods address this issue by identifying what\ndimensions of a single input are most responsible for a DNN's output. The goal\nof this work is to assess the sensitivity of local explanations to DNN\nparameter values. Somewhat surprisingly, we find that DNNs with\nrandomly-initialized weights produce explanations that are both visually and\nquantitatively similar to those produced by DNNs with learned weights. Our\nconjecture is that this phenomenon occurs because these explanations are\ndominated by the lower level features of a DNN, and that a DNN's architecture\nprovides a strong prior which significantly affects the representations learned\nat these lower layers. NOTE: This work is now subsumed by our recent\nmanuscript, Sanity Checks for Saliency Maps (to appear NIPS 2018), where we\nexpand on findings and address concerns raised in Sundararajan et. al. (2018).\n