An unsupervised learning algorithm for multiscale neural activity.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Simultaneous recording of multiscale neural activity (spikes, LFP, ECoG) is now possible.
• Encoding models are crucial for understanding neural mechanisms and developing neurotechnologies like brain-machine interfaces (BMI).
• Existing models struggle to simultaneously capture discrete spikes and continuous LFP/ECoG signals across different timescales.

Purpose of the Study:

• To develop an unsupervised learning algorithm for multiscale state-space encoding models.
• To enable the simultaneous characterization of discrete spike and continuous LFP/ECoG recordings.
• To address the lack of methods for learning complex multiscale neural encoding models from data.

Main Methods:

• Developed a novel expectation-maximization (EM) technique for unsupervised learning.
• Applied the algorithm to learn parameters of multiscale state-space encoding models.
• Utilized simulated multiscale neural data for model training and validation.

Main Results:

• The algorithm accurately learns multiscale state-space encoding model parameters from simulated data.
• The learned model successfully decodes simulated arm movement trajectories from multiscale neural activity.
• Demonstrated the efficacy of the expectation-maximization (EM) technique in this context.

Conclusions:

• The developed unsupervised learning algorithm effectively estimates parameters for multiscale state-space encoding models.
• This approach holds potential for enhancing the performance and robustness of neurotechnologies.
• Facilitates a deeper understanding of neural encoding across multiple signal scales.