iDropout: Leveraging Deep Taylor Decomposition for the Robustness of Deep Neural Networks

摘要

In this work, we present iDropout, a new method to adjust dropout, from purely randomly dropping inputs to dropping inputs based on a mix based on the relevance of the nodes and some randomness. We use Deep Taylor Decomposition to calculate the respective relevance of the inputs and based on this, we give input nodes with a higher relevance a higher probability to be included than input nodes that seem to have less of an impact. The proposed method does not only seem to increase the performance of a Neural Network, but it also seems to make the network more robust to missing data. We evaluated the approach on artificial data with various settings, e.g. noise in data, number of informative features and on real-world datasets from the UCI Machine Learning Repository.