Coarse-to-fine Hierarchical Architecture with Sequential Mamba for Brain Reconstruction
Researchers have developed CHASMBrain, a new two-stage framework for encoding images into fMRI data, inspired by the human visual cortex's hierarchical organization. This architecture uses a dual-stream Mamba design to process global semantic information and local spatial details separately. Experiments on the Natural Scenes Dataset showed CHASMBrain achieved superior performance compared to existing methods, with causal analysis revealing distinct roles for its semantic and spatial processing streams. AI
IMPACT This model offers a new approach to understanding visual representations by mapping them to brain activity, potentially advancing both AI and neuroscience.