Fast Multigroup Gaussian Process Factor Models.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Gaussian processes are vital for dimensionality reduction in neuroscience, modeling high-dimensional neural activity.
• Current Gaussian process factor models struggle with large-scale multipopulation recordings due to cubic runtime scaling.
• Growing neural recording capabilities necessitate more efficient analysis methods.

Purpose of the Study:

• To develop computationally efficient Gaussian process factor models for large-scale multipopulation neural recordings.
• To improve the scalability of analyzing interactions between multiple neural populations.
• To enable advanced analysis techniques to match the pace of modern neuroscience data acquisition.

Main Methods:

• Developed two approximate methods for fitting multigroup Gaussian process factor models: inducing variables and frequency domain approaches.
• Achieved linear scaling with trial length and number of neural groups, a significant improvement over cubic scaling.
• Validated methods through simulations and analysis of neural recordings from hundreds of neurons across multiple brain areas.

Main Results:

• Both approximate methods demonstrated orders of magnitude speed-up in runtime.
• The frequency domain approach offered the most substantial runtime benefits with minimal statistical performance impact.
• Characterized and provided mitigation strategies for estimation biases in the frequency domain method.

Conclusions:

• The developed methods significantly enhance the scalability of Gaussian process factor models for multipopulation neuroscience data.
• These advancements allow for the analysis of larger and more complex neural datasets, facilitating the study of brain function.
• The frequency domain approach is a promising tool for efficient analysis of large-scale neural interactions.