Document Actions

You are here: Home / Bernstein Conference / 2018 / Satellite Workshops / Neuronal Intelligence: Narrowing the gap between neuroscience and AI

Neuronal Intelligence: Narrowing the gap between neuroscience and AI


While advances in deep learning methods have enabled impressive strides in artificial intelligence (AI) and changed our everyday lives, it is still lacks a fundamental feature of biological intelligence: robustness and generalization beyond the immediate data it was trained on. Current models in AI derive their power from the ability to fit almost any arbitrary function. However, the capability for universal approximation is as much a blessing as it is a curse since it is hard to control the behavior of the models outside the domain of training examples the model was exposed to. In stark contrast, most vertebrate brains operate well under extreme changes in signal reliability (e.g. night vs. day) and statistics (e.g. rain forest vs. desert). What are implicit assumptions and computational principles that the brain uses to achieve this level of robustness?

Within the MICrONs program funded by the US government through IARPA, our team set out to expose differences between current state-of-the-art deep learning and the mammalian visual cortex, and to narrow the gap by exploring low-level functional and anatomical patterns in mouse visual cortex and explore their relevance for AI algorithms. In the workshop we present our progress towards this goal through exposing the weaknesses of current deep networks in robustness and generalization on the computational side, and through recording and analyzing the physiology and anatomy of >70,000 neurons from one cubic millimeter of cortex.

Braitenberg’s law of uphill analysis and downhill invention states that it is much easier to understand a complex system through construction than by reverse engineering. The enormous advances in deep learning and neuroscientific recording techniques put the possibility to synergistically advance our understanding of biological and artificial intelligence through a loop of system construction and data analysis within reach. With the proposed workshop we want to discuss current questions on this topic and foster a multi-disciplinary community striving to address them.