Designing a machine that can process information alike and hopefully faster than humans has been a permanent challenging motivation in computing sciences for decades, with von Neumann and Turing discussing brain-like machines as early as the 1950’s. Since the beginning of the information era, the von Neumann architecture has become a standard for such a machine: an electronic board integrates different transistor-based chips in a simple but inefficient way. Comparisons of this architecture to the human brain highlighted significant differences in several aspects: the organizational structure, power requirements, and processing capabilities of the electronic counterpart are far from the performance of the biological equivalent. As a consequence of this trend, since 1990 a new field in Computer Science emerged. Instead of producing software and/or hardware solutions using the classic von Neumann architecture, system designs began to follow the way in which the brain process and store information, referred as Neuromorphic Systems. In spite of recent advancements in Neuromorphic Computing systems such as the IBM True North chip, capable to efficiently process and classify complex signals such as real-time video, there is plenty of room to improve these architectures by adding novel bio-inspiring elements considering the way in which the retina encodes complex visual signals. However, understanding how the nervous system process signals from the physical world, transforming them into a highly efficient code that will determine the behavior of an animal is still an open question in systems neuroscience.
Towards a bio-inspired visual system
Using neurophysiological from our collaborators recorded with Multi-Electrode Array technology, we train artificial neural networks to replicate the retinal responses to different type of stimulus, while keeping the architecture under biological constraints. Through these models we try to understand how images are processed by these networks, and how these biological computation principles can be applied to computer vision, machine learning and other related fields.
