Yongyi Zang

The core of my research centers on “affordances” in neural networks. An affordance refers to all the latent action possibilities of an object (for example, with a push/pull door at a mall entrance, you can either push or pull it—so “push/pull” is the affordance of that door). Extending this idea, in the context of neural networks, I’m interested in those possibilities within a model’s internal representation space that are “manipulable but yet to be discovered.”

My long-term goal is to build a set of tools and methods that allow humans (or other AIs) to discover and create these representation-level affordances, enabling direct control over an AI’s behavioral outputs without retraining the entire model. Concretely, this spans several directions: studying the geometric properties of specific affordances in vector space; performing affordance-based decomposition of learned representations; or building in constraints during training so that certain affordances emerge naturally.

I currently serve as Director of Research and Labs at Smule, where I lead a team uncovering affordances in audio and visual signals and work with the product team to surface these techniques to users in intuitive ways. In my independent research, I focus on affordance discovery for specific network architectures—building behavioral models or imposing spatial constraints to reveal manipulation interfaces that would otherwise remain hidden, so that external systems can interact with these neural networks.

Before this, my background was in music production. These days I sometimes help friends write songs, and occasionally record covers of tracks I love.