Keynotes | K&Ts | GACs | Talks | Posters | Search
Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms
Signatures of discrete action representations emerge in a task-optimized neuro-symbolic model
Carlos G. Correa1, Lucas Y. Tian2, Yue Liu1, Sreejan Kumar3, Daniel Hanuska2, Kedar Garzón Gupta3, Xiao-Jing Wang1, Winrich Freiwald2, Joshua B. Tenenbaum4, Marcelo G Mattar1; 1New York University, 2Rockefeller University, 3Columbia University, 4Massachusetts Institute of Technology
Presenter: Carlos G. Correa
Key to intelligence is the ability to flexibly compose and combine atomic concepts into representations that guide behavior, making "infinite use of finite means". However, modeling often assumes discrete concepts exist and focuses on their _use_, leaving it unclear how they are _learned_ from complex sensory inputs and mapped to motor outputs. To address this limitation, we build on a drawing-like task and behavioral metrics that indicate discrete structure in motor behavior from Tian et al. (2025). Importantly, these discrete concepts were encoded in neural recordings from non-human primates that exhibited these behavioral metrics, validating the metrics. Here, we use this task and behavioral metrics as a modeling target, to quantify how discrete structure arises in a task-optimized neuro-symbolic model (Liang et al., 2022). Our model recapitulates these behavioral metrics, offering a foundation to compare with neural data and understand what drives the emergence of discrete conceptual structure.
Topic Area: Methods, Tools, Theory & Neural Coding