Evaluating and Explaining Generative Adversarial Networks for Continual Learning under Concept Drift

Filip Guzy, Michal Wozniak, Bartosz Krawczyk

Abstract: Generative Adversarial Networks (GANs) are among the most popular contemporary machine learning algorithms. Despite remarkable successes in their developments, existing GANs cannot offer the appropriate tools to monitor their performance in a continual learning scenario when data distribution changes. We propose a complete framework for monitoring and evaluating GANs during the continual learning, explaining their reaction to the data distribution shifts. The proposed approach is the first complete solution for evaluating GANs in drifting environments, additionally adding explainability to the adaptation process. We introduce a novel prequential metric for continual evaluation of GANs. We show how to use various information extracted from streaming GANs to understand the model’s behavior under data changes and gain insight into the nature of concept drift. Our explainable components focus on learning curves under non-stationary data that highlight retaining relevant and forgetting outdated information, as well as of dynamic visualization of changes in relevant regions for drifting images. The proposed tool allows for detecting changes in data, as well as evaluating and explaining the reaction of GANs to the concept shift. Our framework can be downloaded from a public repository https://github.com/w4k2/gan-data-streams.