Abstract:

Biological brains can learn much more autonomously than today’s AI systems and robots. How does this work and can we reproduce such autonomous learning abilities in artificial systems? Over the last years we have been studying this question for the case of visual perception. We have constructed models of how human infants learn to see the world in three dimensions, begin to track moving objects and learn to recognize them without any supervision. We have compared these models to biological data and validated them on physical robots. Studying the computational principles underlying these learning processes, we highlight effective information compression as the central driving force behind our brain’s ability to learn to see without supervision.

>> If you have any questions, please don't hesitate to contact me,