Identification of recurrent dynamics in distributed neural populations

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Data Science

Background:

• Large-scale neural recordings generate complex, high-dimensional time series data.
• Recurrent switching dynamical systems offer a framework for analyzing complex neural data.
• Identifying recurrent linear dynamics in neural time series remains a significant challenge.

Purpose of the Study:

• To develop and test a scalable method for recovering recurrent linear dynamics in neural time series.
• To investigate the ability of time-varying autoregression with low-rank tensors to identify attractor structures.
• To analyze the influence of system parameters and noise on dynamic structure recovery.

Main Methods:

• Application of time-varying autoregression with low-rank tensor decomposition.
• Analysis of stochastic neural mass models with multiple stable attractors.
• Simulations using a human brain connectivity matrix under varying global connection strengths.
• Examination of the impact of forecast time delay on dynamic parameter estimation.

Main Results:

• The proposed method successfully recovers attractor structures in simple dynamical systems via clustering.
• Hierarchical clustering of dynamics was revealed in simulations based on human brain connectivity.
• The study identified the critical role of three key timescales: dynamics, noise, and attractor switching.
• Prediction error minimization alone was shown to be insufficient for meaningful dynamic structure recovery.

Conclusions:

• Scalable low-rank tensor methods can effectively recover recurrent linear dynamics in neural data.
• Understanding the interplay of multiple timescales is essential for accurate dynamic structure inference.
• This approach provides a powerful tool for analyzing complex neural systems and uncovering underlying dynamic principles.