Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching
A new AI framework forecasts short-term neural dynamics, beating established baselines by learning from past brain states.
A team of researchers including Nicole Rogalla, Yuzhen Qin, and Marcel van Gerven has developed a novel AI framework for forecasting brain activity, detailed in a new arXiv paper titled 'Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching.' The model, called Autoregressive Flow Matching (AFM), is a generative AI approach that learns the conditional probability distribution of future neural states. It uses both past brain activity (BOLD signals from fMRI) and concurrent sensory input to predict short-term neural dynamics, explicitly treating brain activity as a temporally evolving process.
The researchers rigorously evaluated AFM using subject-specific models on the Algonauts Project 2025 challenge dataset, a benchmark for predicting brain responses to naturalistic stimuli. The results showed that AFM significantly outperformed two key baselines: the challenge's official General Linear Model (GLM) and a non-autoregressive flow-matching variant. This demonstrates superior generalization and widespread prediction accuracy across the cortex. Ablation studies revealed that access to past BOLD dynamics was the dominant factor driving performance, while the autoregressive structure provided consistent, though modest, gains in short-horizon forecasting.
This work builds on recent advances in transport-based generative modeling, applying them to the complex challenge of neural forecasting. The success of AFM suggests that flow-matching techniques, combined with an autoregressive architecture, are well-suited for modeling the probabilistic and sequential nature of brain activity. The framework's ability to provide probabilistic forecasts—predicting a range of possible future states rather than a single point—is a crucial advancement for applications where uncertainty matters.
- The Autoregressive Flow Matching (AFM) model beat the official Algonauts 2025 challenge GLM baseline in predicting short-term, parcel-wise BOLD activity.
- Ablation studies showed past neural dynamics (BOLD history) were the dominant performance driver, more critical than the autoregressive design itself.
- The framework provides probabilistic forecasts, predicting distributions of future brain states, which is vital for sensitive neurotechnology applications.
Why It Matters
This research provides a more accurate AI model for predicting brain activity, a foundational step toward advanced brain-computer interfaces and personalized neurotherapies.