OmniMouse: Scaling properties of multi-modal, multi-task Brain Models on 150B Neural Tokens

The study introduces OmniMouse, a multi-modal, multi-task model trained on a dataset of 150 billion neural tokens from the visual cortex of mice, enabling flexible test-time regimes for neural prediction, behavioral decoding, and neural forecasting. The model demonstrates state-of-the-art performance while revealing that, unlike traditional AI models, performance gains saturate with increased model size, highlighting the importance of data over parameters in brain modeling. This research suggests potential phase transitions in neural modeling, akin to emergent properties in large language models, which could inform future approaches in AI development.