Neuorescence

Today's AI systems can achieve extraordinary performance, but they still rely heavily on backpropagation and massive training datasets, while biological brains can continuously learn and adapt from real-world experience. We believe one of the keys to this difference may lie in biological credit assignment: how the brain determines which parts of a neural network should change after an action produces a good or bad outcome. Our thesis is that full-network, millisecond-resolution voltage imaging can provide a uniquely direct view of the underlying neural computation during learning, revealing fast distributed dynamics that are difficult to capture with slower or more indirect measurements. We are building the technology required to make these recordings possible and plan to use it to study neural networks as they learn, experimentally testing competing theories of biological credit assignment. Ultimately, if we can identify the computational principles biology uses to learn, we can translate those principles into artificial neural networks and test whether they enable more efficient, continual, real-world learning in AI and robotics.