Abstract:

Although the recently re-discovered field of deep learning has revolutionized areas like computer vision, such approaches have still several limitations that should be kept in mind in order to design more advanced artificial general intelligent systems for human behavior understanding and human-computer interaction, among other complex tasks. In particular, for those two domains, there indeed is a potentially infinite range of input motions with potentially infinite range interpretations.

In this talk, we will cover how the combination of rule-based symbolic reasoning, prior knowledge and common-sense integration, and high-level abstract concepts manipulation can provide more insights to properly interpret the behaviors of humans in front of a camera. If concepts like space, time and object can be represented, the use of hierarchical symbolic systems used for inference can strongly benefit from the so powerful classification performance achieved by neural networks. As a result, by considering deep learning as just part of a more complex and more challenging problem-solving system, a better understanding of human motion can be achieved, including abstraction, structure, intention, open-world and discovery capabilities.