Connections between neuroscience and deep learning
Ascent Robotics is drawing on deep learning to solve robotics and self-driving challenges. The very concept of a neural network is inspired by the brain, and the study of neural networks is punctuated by inspiration from neuroscience. Must we by necessity reverse engineer all of the solutions of the brain to alleviate human labor? Probably not. However, for many tasks that a human would perform, there is still only one example of a system that can do it well - the human brain itself.
Ideas from neuroscience that push the state-of-the-art in deep learning include synaptic plasticity, attention, reinforcement learning, and memory consolidation. Recently, the concept of reasoning has made its way into deep learning. As tasks become more human interpretable, we need methods to evaluate performance. For example, how would we evaluate if a self-driving car “intended” to follow the law, or to avoid a collision? Do we evaluate them with measures designed for the study of the brain and behavior? Critically, our evaluation method will feedback into their design. For these reasons, from low to high-levels, Ascent is drawing parallels between deep learning and neuroscience, and therein finding design solutions where possible.
Anthony did his PhD and postdoctoral fellowship at Columbia University, studying a combination of experimental and theoretical neuroscience, and then computational neuroscience and deep learning at RIKEN in Japan. Recently, he’s been applying a multidisciplinary approach to the challenges at Ascent Robotics.