Interpreting Chest X-Ray Classification Models: Insights and Complexity Measures in Deep Learning

Abstract

Triumph of Deep Learning methods in solving many pressing practical problems in the framework of supervised, un-supervised, self-supervised learning etc. are ubiquitous, yet their interpretation and assessment of robustness in terms of quantifying complexity measure is an ongoing endeavour. This line of work tries to reason this in agreement with empirically justified notion of overparameterization in interpolating regime and implicit regularization of Deep neural network models, in many settings. Our work tries to consolidate these ideas and empirically validate the findings in the setting of supervised Chest X-Ray classification. We find that our design choices in constructing Chest X-Ray classification model progress towards the phenomenon of Double descent and notion of Implicit regularization. We report effect of our designed choices on the classification output quantified in terms of precision, recall and AUC (Area Under ROC Curve) score.