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

Meta-learning In-Context Enables Training-Free Cross Subject Decoding of Vision and Motor

Mu Nan¹, Muquan Yu¹, Margaret M Henderson², Hossein Adeli³, Andrew F. Luo¹; ¹University of Hong Kong, ²Carnegie Mellon University, ³Columbia University

Presenter: Andrew F. Luo

Decoding neural representations across both visual and motor domains is fundamentally constrained by structural and functional inter-subject variability. Current methodologies typically require subject-specific training or gradient-based fine-tuning to account for individualized cortical organization. To address this limitation, we propose a meta-learning in-context framework that enables training-free cross-subject decoding. The model operates through a hierarchical inference process. First, it performs system identification by conditioning on a limited set of paired behavioral and neural data from a target individual, inferring voxel-specific forward encoding models. Second, the architecture aggregates these estimated encoder weights alongside observed neural activations across distributed cortical regions to execute functional inversion. This formulation predicts unseen visual stimuli and motor targets for novel subjects without requiring parametric updates, anatomical coregistration, or shared calibration datasets. By framing neural translation as an in-context learning objective, this approach demonstrates how generalized inverse models can scale across heterogeneous neural datasets and diverse sensorimotor applications.

Topic Area: Methods, Tools, Theory & Neural Coding