Modern artificial intelligence and neuroscience face a common problem: understanding how large distributed nonlinear networks of neurons learn and compute. In both AI and neuroscience we can use common experimental tools involving, observing their input-output behavior, recording the activity of neurons, inspecting the connectivity between them, and causally perturbing them. From these manipulations, how can we reverse engineer neural networks to obtain a conceptual understanding of how computation emerges from them? We will review recent literature in the emerging field of explainable AI that can shed light on these questions, with an eye towards analyzing both biological and artificial networks. Prerequisites: calculus, linear algebra, probability theory essential; basic knowledge of machine learning helpful.
3 units · Letter (ABCD/NP)
Modern artificial intelligence and neuroscience face a common problem: understanding how large distributed nonlinear networks of neurons learn and compute. In both AI and neuroscience we can use common experimental tools involving, observing their input-output behavior, recording the activity of neurons, inspecting the connectivity between them, and causally perturbing them. From these manipulations, how can we reverse engineer neural networks to obtain a conceptual understanding of how computation emerges from them? We will review recent literature in the emerging field of explainable AI that can shed light on these questions, with an eye towards analyzing both biological and artificial networks. Prerequisites: calculus, linear algebra, probability theory essential; basic knowledge of machine learning helpful.
Offered in Autumn 2025 at Stanford University.