Local Explanation Methods for Deep Neural Networks Lack Sensitivity to Parameter Values
This work highlights a critical limitation in interpretability tools for machine learning, potentially affecting researchers and practitioners relying on such explanations for model debugging and trust.
The paper investigates the sensitivity of local explanation methods for deep neural networks to parameter values, finding that explanations from randomly-initialized weights are similar to those from learned weights, suggesting these methods are dominated by lower-level features and architectural priors.
Explaining the output of a complicated machine learning model like a deep neural network (DNN) is a central challenge in machine learning. Several proposed local explanation methods address this issue by identifying what dimensions of a single input are most responsible for a DNN's output. The goal of this work is to assess the sensitivity of local explanations to DNN parameter values. Somewhat surprisingly, we find that DNNs with randomly-initialized weights produce explanations that are both visually and quantitatively similar to those produced by DNNs with learned weights. Our conjecture is that this phenomenon occurs because these explanations are dominated by the lower level features of a DNN, and that a DNN's architecture provides a strong prior which significantly affects the representations learned at these lower layers. NOTE: This work is now subsumed by our recent manuscript, Sanity Checks for Saliency Maps (to appear NIPS 2018), where we expand on findings and address concerns raised in Sundararajan et. al. (2018).