Unsupervised learning of multiscale switching dynamical system models from multimodal neural data.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Neural population activity often shows switching dynamics, indicating non-stationarity.
• Existing models struggle with multimodal neural data and lack of regime labels.
• Accurate models are crucial for understanding behavior encoded in neural activity.

Purpose of the Study:

• Develop an unsupervised learning algorithm for switching dynamical system models.
• Enable learning from multiscale neural data without requiring regime labels.
• Fuse information from multiple neural modalities to capture complex dynamics.

Main Methods:

• Developed a novel unsupervised learning algorithm.
• Utilized multiscale neural observations (e.g., spike and local field potentials).
• Algorithm learns parameters of switching multiscale dynamical system models.

Main Results:

• The switching multiscale dynamical system models achieved more accurate behavior decoding than single-scale models.
• Multiscale neural fusion was successfully demonstrated.
• Models outperformed stationary multiscale models, highlighting the importance of regime switches.

Conclusions:

• The unsupervised framework improves modeling of multiscale neural dynamics using multimodal recordings.
• This approach enhances brain-computer interface performance and robustness.
• The method advances understanding of the neural basis of behavior.