Dynamic Compression Flows for Neuroscience Data.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Data Science

Background:

• Neuroscience studies reveal large neuronal populations in simple behaviors, yet neural activity dynamics are often low-dimensional.
• Current methods for low-dimensional latent representations of time series data in neuroscience often neglect temporal structure or use latent dynamical systems.
• These existing approaches can distort temporal dynamics or lead to ambiguous latent representations.

Purpose of the Study:

• To develop a novel method for identifying identifiable low-dimensional representations of neural time series data.
• To preserve temporal relations within the data during dimensionality reduction.
• To improve the interpretability of neural dynamics and the quality of data reconstructions.

Main Methods:

• Introduced a novel flow-matching approach using a pair of flow fields.
• One flow field governs time evolution, while the other maps data to a low-dimensional latent space.
• The dimension-reducing flow is trained to minimize temporal dynamic distortions, incorporating soft constraints for low-dimensional support.

Main Results:

• The dual flow approach produces more interpretable dynamics compared to existing methods.
• Achieved higher-quality reconstructions of both neural and behavioral data.
• Demonstrated superior performance in noise-dominated datasets where conventional methods fail.

Conclusions:

• The proposed dual flow method effectively learns identifiable low-dimensional representations that preserve temporal dynamics.
• This approach offers significant advantages in interpretability and reconstruction quality for neuroscience time series data.
• The method shows promise for analyzing complex, noisy neural and behavioral data effectively.