Detecting change points in neural population activity with contrastive metric learning

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Change point detection is crucial for analyzing neural data, but existing methods struggle with complex free behaviors and sparse neural signals.
• Identifying shifts in neural response distributions is challenging due to diverse behavioral states and varying change rates.
• High-dimensional neural recordings often exhibit sparse changes, leading to potential false positives with conventional change point detection techniques.

Purpose of the Study:

• To develop a novel approach for detecting changes in neural population states during free behavior.
• To address limitations of existing methods in handling diverse neural activity and arousal states.
• To improve the accuracy and interpretability of change point detection in complex neural recordings.

Main Methods:

• A contrastive learning framework is employed to learn a metric for change point detection.
• The model maximizes Sinkhorn divergences of neuron firing rates across labeled change points.
• The method is applied to a 12-hour neural recording from a freely behaving mouse.

Main Results:

• The proposed method successfully detects changes in neural population states corresponding to sleep stages and behavior.
• Learning a metric improves the detection of change points compared to existing approaches.
• The analysis provides insights into specific neurons and neuronal subgroups involved in detecting different types of neural switches.

Conclusions:

• The contrastive learning approach offers a robust method for change point detection in complex neural data.
• This technique enhances our ability to understand neural dynamics during naturalistic behaviors.
• The findings contribute to more refined neural data analysis pipelines for neuroscience research.