Using noise resilience for ranking generalization of deep neural networks
This work addresses the problem of predicting generalization error for deep neural networks, which is crucial for understanding why these models generalize well on real-world data despite their capacity to fit random labels.
This paper proposes several measures to predict the generalization error of a deep neural network given its training data and parameters. One measure, based on noise resilience, achieved 5th position in the PGDL competition at NeurIPS 2020.
Recent papers have shown that sufficiently overparameterized neural networks can perfectly fit even random labels. Thus, it is crucial to understand the underlying reason behind the generalization performance of a network on real-world data. In this work, we propose several measures to predict the generalization error of a network given the training data and its parameters. Using one of these measures, based on noise resilience of the network, we secured 5th position in the predicting generalization in deep learning (PGDL) competition at NeurIPS 2020.