Point process models for sequence detection in high-dimensional neural spike trains.

Area of Science:

• Neuroscience
• Statistical Neuroscience
• Computational Neuroscience

Background:

• Sparse neural spike sequences are fundamental to cognitive functions like working memory, motor control, and learning.
• Unsupervised discovery of these sequences is a significant challenge in statistical neuroscience.
• Previous models, like convolutive nonnegative matrix factorization, have limitations including data discretization and suboptimal criteria.

Purpose of the Study:

• To address limitations of existing models for discovering sparse neural spike sequences.
• To develop a more accurate and flexible model for analyzing neural spike trains.
• To introduce a novel point process framework for ultra-sparse sequence event representation.

Main Methods:

• Developed a point process model operating on continuous time, analyzing individual spikes.
• Represented sequence occurrences as marked events in continuous time for ultra-sparse modeling.
• Introduced learnable time warping parameters to capture sequences of varying durations.

Main Results:

• The new point process model overcomes limitations of discretization and least-squares criteria.
• The model provides uncertainty estimates for predictions and parameters.
• Demonstrated effectiveness on experimental neural recordings from songbirds and rodents.

Conclusions:

• The developed point process model offers a powerful new approach for analyzing fine-scale neural sequences.
• Learnable time warping enables modeling of dynamic sequence durations observed in neural circuits.
• This framework advances spike train modeling and understanding of neural coding.