Informativeness of Degraded Data in Training a Classification System

Abstract

Many recent solutions have been proposed to mitigate the vulnerability of machine learning models when they are subject to limited or degraded data. However, the effects of using degraded data for purposes of training or testing a classification system are not fundamentally studied. In this work, we propose a methodology for studying the effects of degradations (due to additive noise, compression artifacts, and blur) that is based on the active learning framework for studying the informativeness of data samples. We provide experimental results using the action recognition video dataset UCF101 to validate its utility. We shed light on the importance of studying the effects of degraded data by showing to which extent degraded samples can be more informative than unedited high quality samples in training a classification system.