| name | Autoregressive Flow Matching for Neural Dynamics |
| description | Framework for probabilistic prediction of neural population dynamics using autoregressive flow matching (AFM). Addresses the inherent stochasticity and nonlinearity of neural activity by leveraging transport-based generative modeling to forecast neural responses from multimodal sensory input at scale.
|
Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching
Paper: arXiv:2604.11178
Authors: Elena B. Trifonova, Manuel S. Hafliger, Leo K. K. Ma, Arnu V. K. Tubaon, Gilad A. Silberberg
Categories: q-bio.NC, cs.LG
Year: 2025
Overview
Forecasting neural activity in response to naturalistic stimuli remains a key challenge for understanding brain dynamics and enabling downstream neurotechnological applications. This paper introduces a generative forecasting framework for modeling neural dynamics based on autoregressive flow matching (AFM). Building on recent advances in transport-based generative modeling, the approach probabilistically predicts neural responses at scale from multimodal sensory input.
The framework addresses two fundamental properties of neural population activity:
- Inherent stochasticity — neural responses to identical stimuli vary across trials
- Nonlinearity — neural dynamics exhibit complex, nonlinear temporal dependencies
By combining flow matching with autoregressive structure, AFM provides full conditional probability density estimates rather than just point predictions, enabling calibrated uncertainty quantification for neural forecasting.
Key Concepts
Flow Matching
- A transport-based generative modeling paradigm that learns deterministic ODE-based transformations (vector fields) to map a simple source distribution (e.g., Gaussian noise) to a complex target distribution
- The training objective minimizes the difference between the learned vector field and an optimal transport vector field, yielding efficient and stable training
- Unlike diffusion models, flow matching uses straight-line interpolation paths between source and target, reducing sampling cost
Autoregressive Factorization
- Decomposes complex high-dimensional joint distributions into a sequence of context-dependent conditional flows
- Each prediction step conditions on previously generated outputs, propagating temporal context forward
- Enables sequential generation of future neural activity predictions conditioned on observed history
Probabilistic Neural Forecasting
- Rather than producing single deterministic predictions, AFM generates full probability distributions over future neural states
- Enables uncertainty quantification, calibration analysis, and sampling of diverse plausible trajectories
- Critical for closed-loop neurotechnology applications where prediction confidence matters
Methodology
Architecture
- Transformer Encoder — processes historical neural activity data and sensory stimuli to build a rich contextual representation
- Conditional Flow Model — uses autoregressive structure to generate future neural activity predictions conditioned on the encoded context
- The learned vector field defines an ODE that transforms noise into predicted neural states over a continuous time parameter
Training Objective
- Derived from the flow matching principle: minimize the discrepancy between the learned conditional vector field and an optimal transport vector field connecting source and target distributions
- Autoregressive factorization allows the model to decompose the joint predictive distribution into tractable conditional distributions
- The model learns to predict neural population responses conditioned on multimodal sensory input (e.g., auditory, visual stimuli)
Inference
- At prediction time, noise samples are drawn from the source distribution and transported through the learned ODE to produce samples from the predicted neural distribution
- Multiple samples can be drawn to characterize prediction uncertainty and generate diverse plausible neural trajectories
- The autoregressive rollout generates sequential future time steps
Applications
- Closed-loop neurotechnology — real-time prediction of neural states for brain-computer interfaces (BCIs) with uncertainty-aware control
- Neural dynamics modeling — understanding how neural populations encode and process naturalistic sensory stimuli
- Neuroscience research — probabilistic forecasting of neural responses to aid experimental design and hypothesis testing
- Brain-machine interfaces — improving decoding accuracy through calibrated probabilistic predictions
- Neural simulation — generating realistic neural population activity patterns for computational studies
Key Insights
-
Generative over discriminative — By modeling the full conditional distribution rather than just point estimates, AFM captures the inherent trial-to-trial variability of neural responses, which deterministic models miss.
-
Transport-based efficiency — Flow matching offers computational advantages over diffusion-based approaches (fewer sampling steps, straighter paths), making it more suitable for real-time or near-real-time neurotechnology applications.
-
Autoregressive structure for temporal coherence — The autoregressive factorization naturally captures the temporal dependencies in neural dynamics, producing coherent multi-step forecasts.
-
Scalability — The framework is designed to handle neural population activity at scale (large numbers of simultaneously recorded neurons), addressing a key limitation of previous probabilistic approaches.
-
Multimodal sensory conditioning — AFM can condition predictions on rich, multimodal sensory input streams, enabling forecasting of neural responses to complex naturalistic stimuli.
-
Short-term forecasting strength — The method is particularly effective for short-term probabilistic forecasting of neural dynamics, positioning it as a practical tool for closed-loop applications where near-term prediction accuracy is paramount.
References
- Trifonova, E. B., Hafliger, M. S., Ma, L. K. K., Tubaon, A. V. K., & Silberberg, G. (2025). Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching. arXiv:2604.11178.
- Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., & Le, M. (2023). Flow Matching for Generative Modeling. ICLR 2023.
- Xie, A., & Stojanov, S., et al. (2024). Autoregressive Flow Matching for Motion Prediction. Related work extending AFM to sequential continuous data.
- Albergo, M. S., & Vanden-Eijnden, E. (2023). Building Normalizing Flows with Stochastic Interpolants. ICLR 2023.
- Related Semantic Scholar entry: https://www.semanticscholar.org/paper/99ca1ab8215462be7be7aab1560b3919dc3ebe77