With deep learning models performing exceptionally well across a wide range of tasks, it is more important than ever to understand how they work. Understanding these models is important to improve safety and reliability, and to place appropriate trust in them. In this talk, I will focus on a model-level approach of understanding neural networks through individual neurons or small units in the network inspired by neuroscience. In particular, I will discuss a series of our papers developing scalable and efficient automated interpretability techniques for neural networks, which automates the labor intensive process of understanding individual neurons across deep vision models and large language models. These research efforts allow us to scale neuron-level understanding up to large models with thousands or millions of individual components and are 10-200x faster in vision models, and 2-3x more efficient in language models than existing methods. 

"Learning to Move, Learning to Play, Learning to Animate" is a cross-disciplinary multimedia performance that challenges the human-centric perspective, offering a new way to experience the world. The artwork is a joint effort from PhD students from Music, Visual Arts, and Computer Science departments. This performance features robots made from natural materials, using technologies like AI-generated visuals, real-time bio-feedback, and electroacoustic sound.Inspired by ecologist David Abram’s concept of the “more-than-human world,” our work explores intelligences beyond the human, questioning whether intelligence is exclusive to humans or can exist in wood, stone, metal, and silicon.