Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

SwiFT V2: Towards Large-scale Foundation Model for Functional MRI

Jubin Choi¹, Heehwan Wang¹, Junbeom Kwon¹, Shinjae Yoo², Jiook Cha¹; ¹Seoul National University, ²Brookhaven National Lab

Presenter: Jiook Cha

Foundation models, leveraging large-scale datasets and extensive parameter counts, show unprecedented capabilities across various domains. Recent studies have explored foundation models for neuroimaging to effectively capture the complex dynamics of the human brain. However, training such models end-to-end on four-dimensional functional MRI data remains unexplored. Here, we introduce SwiFT V2, a fMRI foundation model based on the 4D Swin fMRI Transformer. We pre-trained SwiFT V2 using masked image modeling on large-scale aggregated resting-state fMRI datasets of 49,321 subjects. Especially, we trained models up to 8.8 billion parameters with maximal update parameterization technique, leading to stable and efficient scaling. We observed that these models follow neural scaling laws, where performance predictably improves with scale. Also, we showed that masked modeling pre-training enhances performance across various downstream tasks. These results validate the application of scaling principles to fMRI modeling and motivate the further development of large foundation models for neuroscience.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF