Notice: Function _load_textdomain_just_in_time was called incorrectly. Translation loading for the updraftplus domain was triggered too early. This is usually an indicator for some code in the plugin or theme running too early. Translations should be loaded at the init action or later. Please see Debugging in WordPress for more information. (This message was added in version 6.7.0.) in /home/dh_jgw4fm/euvip2024.org/wp-includes/functions.php on line 6114
Slava Voloshynovskiy – EUVIP 2024

Slava Voloshynovskiy

University of Geneva

Vision Through the Information-Theoretic AI Lens: Variational and Contrastive Techniques in Explainable AI

Abstract

In this keynote, we delve into the transformative impact of the information-theoretic framework on enhancing explainability in the field of computer vision, with a special focus on its application in autoencoders, self-supervised learning systems, and generative models. Central to our exploration is the decomposition of mutual information into variational and contrastive components, a methodology that not only deepens our understanding of AI systems but also facilitates the development of more transparent and interpretable models in vision applications. This approach is particularly crucial in improving the analysis and design of existing autoencoder architectures, modern self-supervised learning, and generative models, aligning them more closely with the principles of explainable AI. The presentation will highlight the significance of designing advanced, explainable sampling schemes, especially in specialized areas such as astronomy and medical imaging. These schemes play a vital role in enhancing the accuracy and effectiveness of applications in these fields while ensuring their operational mechanisms are clear and comprehensible. Furthermore, the talk will cover the strategy of conducting data analysis directly in the UV space or k-space, and the Fourier domain. This approach not only optimizes tasks like classification, source estimation, anomaly detection, and disease identification but also adds a layer of transparency to these processes. Marrying complex theoretical constructs with their practical applications will contribute to the advancement of explainable, reliable, and effective technology across diverse fields, ranging from astronomy to medical imaging.

Speaker’s Bio

Slava Voloshynovskiy (IEEE SM’11), a Professor at the University of Geneva’s Department of Computer Science, leads the Stochastic Information Processing group. He earned his Ph.D. in Electrical Engineering from State University Lvivska Polytechnika, Ukraine, after completing his Radio Engineer degree at Lviv Polytechnic Institute. His research focuses on image processing, multimedia security, privacy, and machine learning. Voloshynovskiy has published over 350 journal and conference papers and holds twelve patents. He served as an Associate and Senior Editor for IEEE journals and was an active member of various IEEE committees. Earlier in his career, he was a visiting scholar at the University of Illinois at Urbana-Champaign. In addition to his academic role, he has experience as a consultant in multimedia security and co-founded several companies specializing in copyright and brand protection. He was also the recipient of the Swiss National Science Foundation Professorship Grant in 2003.