Benchmarking Probabilistic Time Series Forecasting Models on Neural Activity.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Neural activity forecasting is crucial for understanding brain function and developing closed-loop systems.
• Deep learning excels in time series forecasting but is underutilized for neural data.

Purpose of the Study:

• To systematically evaluate deep learning models for neural activity forecasting.
• To compare their performance against classical statistical models.

Main Methods:

• Evaluated eight probabilistic deep learning models, including foundation models, on mouse cortical activity data.
• Utilized wide-field imaging for spontaneous neural activity recordings.
• Compared deep learning models against four classical statistical models and two baselines.

Main Results:

• Several deep learning models consistently outperformed classical approaches across various prediction horizons.
• The top-performing model achieved accurate forecasts up to 1.5 seconds into the future.
• Demonstrated the potential of deep learning for predicting complex neural dynamics.

Conclusions:

• Deep learning models offer a powerful tool for advancing neural activity forecasting.
• These findings support future applications in brain-computer interfaces and neural control.
• Opens new research avenues for exploring the temporal structure of neural activity.