Deep learning algorithms have led to unprecedented improvement of computer
vision, natural language processing and other fields of artificial
intelligence.
These algorithms are so resilient that sometimes one can produce useful
performance even by considering them as a black box. Still, it usually pays off
to promote desired behaviour through specific recognition modules. These
modules
can encourage features of varying complexity, increase the receptive field,
raise awareness of the unknown or allow better exploitation of training data. I
will present some of the concepts that have helped us to exceed the state of
the
art in several branches of dense semantics and associate them with the
respective kinds of inductive bias.


Bio: Siniša Šegvić has received a PhD degree in 2004 at the University of
Zagreb Faculty of Electrical Engineering and Computing (UniZg-FER). He was a
postdoc researcher at IRISA, Rennes and at TU Graz. He has worked at UniZg-FER
as a computer science professor since 2008. His research and professional
interests include computer vision, visual recognition, scene understanding, and
dense prediction with deep convolutional models. He has led several academic
and
industrial projects in the field of computer vision at the national level. He
participated in industrial development as a technical consultant. He advises
several PhD students funded by national projects, EU projects, and private
companies. His research group has achieved notable results at competitions such
as Cityscapes, WildDash, Robust vision challenge, Fishyscapes, Segment Me If
You
Can, and ACDC.


Prof. Segvic is visiting VRG and will stay for 4 days from Monday to Thursday.
He will be seated at G10.