Classical notions of information, such as Shannon entropy, measure the amount of information in a signal in terms of the frequency of occurrence of symbols. Such notions are very useful for tasks such as data compression. However, they are not as useful for tasks such as visual recognition, where the semantic content of the scene is essential. Moreover, classical notions of information are affected by nuisance factors, such as viewpoint and illumination conditions, which are irrelevant to the recognition task. The goal of this workshop is to bring together researchers in computer vision, machine learning and information theory, to discuss recent progress on defining and computing new notions of information that capture the semantic content of multi-modal data. Topics of interest include but are not limited to information theoretic approaches to scene understanding, representation learning, domain adaptation, and generative adversarial networks as well as the interplay between information and semantic content.
- René Vidal, Seder Professor of Biomedical Engineering and Director of the Mathematical Institute for Data Science, Johns Hopkins University.
- Michael Jordan, Pehong Chen Distinguished Professor of Electrical Engineering and Computer Science, and Statistics, UC Berkeley.
- Emmanuel Candes, Barnum-Simons Chair and Professor of Mathematics and Statistics, Stanford University.
- Donald Geman, Johns Hopkins University
- Stefano Soatto, UCLA
- Rama Chellapa, UMD
- John Shawe-Taylor, UCL
- Jason Lee, USC
- Mark Girolami, Imperial College London
- Arnaud Doucet, Oxford University
- Josef Kittler, Surrey University
- 9:00 AM: Donald Geman, "Semantic Information Pursuit"
- 9:45 AM: Stefano Soatto, "Information Dropout"
- 11:00 AM: Rama Chellapa, "Information-Driven Domain Adaptation"
- 11:45 AM: John Shawe-Taylor, "Generalization Bounds for Deep Learning"
- 2:00 PM: Jason Lee, "Gradient Based Optimization for GANs"
- 2:45 PM: Mark Girolami, "Quantifying Uncertainty"
- 4:00 PM: Arnaud Doucet, "Hamiltonian Variational Autoencoders"
- 4:45 PM: Josef Kittler, "Information Fusion"