Physics Colloquium: Prof. Dezhe Jin

Abstract: AI has achieved or surpassed human-level performance in many tasks, but current artificial neural networks require substantial computational and energy resources. These costs increasingly constrain the scaling of AI models and their deployment on edge devices such as robots. In contrast, the brain performs sophisticated real-time computations with remarkable energy efficiency. Understanding the computational principles that enable this efficiency may therefore provide new directions for energy-efficient AI.

Neurons in the brain communicate primarily through spikes, which are nonlinear excursions of membrane potentials. Because generating and transmitting spikes is energetically costly, the brain often operates with remarkably sparse spiking activity. This raises a fundamental question: how much useful computation can be achieved with very few spikes?

In this talk, I will show that important computations, including finding the maximum among competing inputs and recognizing auditory objects in continuous sensory streams, can be performed using ultra-sparse spikes. I will also discuss how these computational principles can be implemented in neuromorphic hardware, including spin-based systems, providing a potential route toward more energy-efficient AI.