WiCAT Model Enables Zero-Shot Brain Decoding Across Subjects
Self-supervised model decodes behavior from brain imaging without training on new subjects.
Widefield calcium imaging captures brain-wide cortical dynamics at unprecedented scale, but its high dimensionality and task-irrelevant activity have limited modeling to single-session analyses. Researchers at the University of Southern California (Mohammad Hosseini, Eray Erturk, Saba Hashemi, Maryam M. Shanechi) propose WiCAT—the first multi-subject foundation model for this modality. WiCAT introduces atlas-grounded spatiotemporal tokenization that eliminates session-specific components, then uses self-supervised pretraining to learn globally shared representations across subjects.
The model outperforms baseline single-session models across multiple widefield datasets, enabling lightweight downstream decoders that transfer across subjects, tasks, and even datasets. Crucially, WiCAT achieves robust zero-shot continuous behavior decoding on unseen subjects—meaning it can predict behavior without any fine-tuning—and can reconstruct activity in left-out brain regions. This breakthrough, published at ICML 2026, marks a significant step toward scalable, subject-invariant neural decoding and foundation modeling for neuroscience.
- WiCAT uses atlas-aligned spatiotemporal tokenization to model brain-wide cortical dynamics across subjects.
- Achieves zero-shot behavior decoding on unseen subjects without additional training or session-specific components.
- Transfers across subjects, tasks, and datasets, outperforming single-session models on multiple widefield calcium imaging datasets.
Why It Matters
Enables scalable brain decoding without subject-specific training, advancing neuroscience and brain-computer interfaces.