Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning

Background:

• High-dimensional neural activity is challenging to interpret.
• Current methods use a two-stage approach: smoothing then dimensionality reduction.
• These methods lack principled ways to select smoothing parameters and account for spiking variability.

Purpose of the Study:

• To develop improved methods for extracting smooth, low-dimensional neural trajectories.
• To unify smoothing and dimensionality reduction into a single probabilistic framework.
• To apply these methods to neural recordings from macaque motor cortex during reach tasks.

Main Methods:

• Extensions to two-stage methods for principled smoothing and variability accounting.
• Development of Gaussian-process factor analysis (GPFA), a novel unified probabilistic approach.
• Application to 61-neuron simultaneous recordings in macaque premotor and motor cortices.

Main Results:

• Extended two-stage methods showed improved predictive ability over standard approaches.
• GPFA further enhanced predictive ability, outperforming two-stage methods by tens of percent in simulations.
• Observed direct evidence of neural state convergence during motor planning.

Conclusions:

• GPFA offers a powerful, unified framework for neural trajectory extraction.
• The methods provide better insights into neural circuit dynamics and single-trial neural population activity.
• GPFA significantly advances the analysis of complex neural data.