Multimodal subspace identification for modeling discrete-continuous spiking and field potential population activity

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Learning dynamical latent state models from multimodal neural data (spiking and field potentials) is crucial for understanding brain dynamics and decoding behavior.
• Efficient unsupervised learning methods are needed for real-time applications like brain-machine interfaces (BMIs), but multimodal spike-field data present challenges due to heterogeneous distributions and timescales.
• Existing methods struggle with computational efficiency for complex multimodal neural data.

Purpose of the Study:

• To develop a computationally efficient unsupervised learning method for modeling and dimensionality reduction of multimodal spike-field data.
• To enable accurate real-time learning for applications such as brain-machine interfaces.
• To improve the understanding of collective low-dimensional dynamics in neural activity.

Main Methods:

• Developed a multiscale subspace identification (multiscale SID) algorithm for computationally efficient learning.
• Derived a new analytical subspace identification (SID) method for combined Poisson and Gaussian observations (spike-field activity).
• Introduced a constrained optimization approach for learning valid noise statistics, crucial for multimodal inference.

Main Results:

• Multiscale SID accurately learned dynamical models and extracted low-dimensional dynamics from multimodal spike-field signals.
• The method effectively fused multimodal information, outperforming single-modality approaches in identifying dynamical modes and predicting behavior.
• Multiscale SID demonstrated significantly lower training times compared to existing expectation-maximization methods, with comparable or superior accuracy.

Conclusions:

• Multiscale SID is an accurate and efficient method for modeling multimodal neural data.
• The algorithm is particularly beneficial for applications requiring efficient learning, such as online adaptive BMIs and reducing offline analysis time.
• This method advances the ability to track non-stationary neural dynamics and analyze complex brain activity.