Siniša Šegvić, Faculty of Electrical Engineering and Computing, University of Zagreb

Elements of Learning Algorithms for Natural Scene Understanding

Abstract

Deep learning has led to unprecedented improvement of computer vision, natural language processing and other fields of artificial intelligence. However, our models still underperform on unusual and adversarial test examples, while offering limited interpretability and explainability. Nevertheless, experienced practitioners seldom regard their models as black boxes. Instead, they promote desired behaviour through suitable kinds of inductive bias and careful exploitation of available data. I will illustrate these concepts by describing elements of learning algorithms which have been extensively exploited within my research group in the past few years.

The second part of my talk will describe our ongoing collaborations with the local industry. I will point out advantages of such arrangements for all involved parties. The talk will conclude with a brief overview of current challenges and opportunities in our field.

Short Biography

Siniša Šegvić was a postdoc researcher at IRISA, Rennes and at TU Graz. He led three research projects of the Croatian Science Foundation (MultiCLOD, MASTIF, ADEPT) as well as several industrial research projects funded by local companies. He has participated in the research center of excellence DataCross, and several ERDF projects (SafeTram, A-UNIT). His research and professional interests include computer vision, visual recognition, dense semantic prediction and forecasting, as well as generative modelling with normalizing flows. He has published several papers at top conferences on computer vision and artificial intelligence. He has participated in industrial development as a technical consultant. He advises several full-time PhD students funded by EU projects, national projects and private companies. His research group has achieved notable results while participating at computer vision challenges such as WildDash, Robust vision challenge, Fishyscapes and Cityscapes.