METHODS FOR EVALUATING INTERPRETATIONS OF COMPUTER VISION MODELS

Kapustin Ilia Sergeevich, Romashkova Oxana Nikolaevna

The article examines the problem of interpretability of computer vision models. The author notes that although there are many methods designed to explain the decisions made by deep neural networks, little effort has been made to ensure that these explanations are objectively relevant. The article proves the correlation between the quality of a model and the quality of its interpretation, considers metrics for assessing interpretations taken from the field of algorithmic stability: average generalization (MeGe) and relative consistency (ReCo), and also proposes a new metric for assessing the quality of interpretations of computer vision models, taking into account quality of the original model.

Keywords:interpretation, machine learning, correlation, MeGe, ReCo, interpretation quality, model evaluation