Computer Vision News - March 2022

10 ICLR 2022 Workshop the world will enable agents to have more robust behavior and to generalize better, but they focus on different things. Causality assumes structure and then shows how robust an agent can be. In contrast, object-centric learning starts from an unstructured input and aims to infer a useful decomposition into meaningful factors. Both fields will be represented at the Objects, Structure, and Causality workshop, whichWilka hopes will help foster collaboration. “ You can make good inferences about causality if you know the causal factors in the world, and one important causal factor is an object, ” he explains. “ We are getting to a point where we can at least infer objects in simple scenes, but there’s no state of the art that merges these. In causality, nobody knows how to go from what you see with your eyes to the causal factors. We hope we can get an end-to-end system that can discover the causal factors and then discover causality on top of that . That will be an important part of future AI. ” Outside of this, Wilka works in deep reinforcement learning, which he laughs makes him “ a bit of an outsider ” as one of the workshop organizers. His work involves developing networks that choose actions in virtual environments. Currently, he is working on a network that tries to find objects in a virtual kitchen. Part of that is understanding that objects have shared relations across experiences. Wilka Carvalho is a PhD candidate in computer science at the University of Michigan, under the supervision of Satinder Singh and Honglak Lee. As a co-organizer of the first Objects, Structure, and Causality workshop at this year’s Virtual ICLR, he is here to tell us more about the topic and what we can expect from the event in April. Object-centric representation learning and causal machine learning have similar goals, but the two fields have been working independently. Both are motivated by the idea that structured representations of ELEMENTS OF REASONING: OBJECTS, STRUCTURE, AND CAUSALITY – VIRTUAL ICLR 2022 WORKSHOP