Visual Intelligence research presented in invited keynote talks and tutorials at the top venues

Visual Intelligence’s advances in fundamental and applied deep learning research is gaining widespread attention and visibility.

In the summer of 2023, Visual Intelligence research was presented in two of the top venues within the field.

At the 2023 IEEE International Conference on Speech, Acoustics and Signal Processing (ICASSP), Visual Intelligence director Robert Jenssen gave a 3 hour tutorial on “Information Theory meets Deep Learning”. He gave the tutorial together with Professor Jose Principe, University of Florida, and Associate Professor Shujian Yu, Vriie Universiteit Amsterdam and UiT The Arctic University of Norway.

Director Jenssen also gave one of the invited keynote talks of the workshop “DEEP LEARNING: Theory, Algorithms and Applications” organized in Trento, Italy, June 21-23, 2023. The workshop was organized by Pierre Baldi, University of California, Irvine, USA, Tomaso Poggio, MIT, Boston, USA, Marc'Aurelio Ranzato, DeepMind, London, UK, Alessandro Sperduti, University of Padova, Italy, and Paolo Traverso, Fondazione Bruno Kessler, Trento, Italy.

Jenssen’s talk was entitled “On representation learning with information theoretic criteria and a new method for representation learning interpretability”. The main emphasis was on Visual Intelligence’s recently published method for interpreting features in input images to so-called representation learning systems. The new method, named “Relax” was published in the International Journal of Computer Vision in 2023 as a first of its kind, and has already gained attention and follow-up work from researchers in the wider deep learning community.

Of all the tutorials at ICASSP 2023, this tutorial had the highest number of attendees. The tutorial was composed of three parts. In part I, Principe laid out the fundamentals of statistical properties of information extraction focusing in particular on measures of entropy, divergence, and mutual information. In part II, Yu explained new information theoretic learning criteria developed by the team of presenters in the recent years. Finally, in part II, Jenssen showed examples of deep learning tasks, such as e.g. multi-view clustering, where the information theoretic approach to learning has been pursued.

Director Jenssen also gave one of the invited keynote talks of the workshop “DEEP LEARNING: Theory, Algorithms and Applications” organized in Trento, Italy, June 21-23, 2023. The workshop was organized by Pierre Baldi, University of California, Irvine, USA, Tomaso Poggio, MIT, Boston, USA, Marc'Aurelio Ranzato, DeepMind, London, UK, Alessandro Sperduti, University of Padova, Italy, and Paolo Traverso, Fondazione Bruno Kessler, Trento, Italy.

Jenssen’s talk was entitled “On representation learning with information theoretic criteria and a new method for representation learning interpretability”. The main emphasis was on Visual Intelligence’s recently published method for interpreting features in input images to so-called representation learning systems. The new method, named “Relax” was published in the International Journal of Computer Vision in 2023 as a first of its kind, and has already gained attention and follow-up work from researchers in the wider deep learning community.

‍