The Computational Power of Single Neurons: Dendrites as Architectural Primitives for Neural Networks

Nearly all of AI, and most models of the brain, treat neurons as simple units that sum their inputs, but real cortical neurons have elaborate dendritic trees with nonlinear channels and thousands of synapses, which suggests rich computation within a single cell. By training realistic, detailed neuron models with gradient descent, I show that a single cortical neuron can solve sensory tasks such as visual and auditory classification, as well as abstract nonlinear parity tasks, far beyond the reach of point neurons. These results motivate my current and future work, which scales this approach to networks of detailed neurons and treats dendrites as architectural primitives for understanding brain circuits and building more powerful, energy-efficient AI.