Representation Learning for Dynamic Functional Connectivities via Variational Dynamic Graph Latent Variable Models.

Area of Science:

• Computational neuroscience
• Machine learning
• Network neuroscience

Background:

• Latent variable models (LVMs) analyze neural population spikes but lack neurophysiological interpretation.
• Dynamic functional connectivities (DFC) are increasingly linked to cognitive and behavioral neural activity patterns.
• Existing LVMs struggle to model evolving relationships in neural data.

Purpose of the Study:

• To associate low-dimensional neural activity structures with dynamic functional connectivities.
• To develop a novel representation learning model for neural data analysis.
• To improve the interpretability of latent dynamics in neural population activity.

Main Methods:

• Introduced a Variational Dynamic Graph Latent Variable Model (VDGLVM).
• Utilized a dynamic graph as the latent variable within a variational information bottleneck framework.
• Employed a graph generative model and graph neural network to capture unobserved node communication.

Main Results:

• The VDGLVM model demonstrated guaranteed behavior-decoding performance.
• Successfully associated inferred latent dynamics with probable dynamic functional connectivities (DFC).
• Improved upon existing LVMs by providing a link to neurophysiological interpretations.

Conclusions:

• The VDGLVM offers a powerful framework for understanding the relationship between neural dynamics and brain connectivity.
• This approach enhances the interpretability of latent variable models in neuroscience.
• The model advances the analysis of neural population data by integrating dynamic graph structures.