Real-Time Point Process Filter for Multidimensional Decoding Problems Using Mixture Models

Area of Science:

• Computational Neuroscience
• Machine Learning
• Signal Processing

Background:

• Real-time neural decoding is crucial for closed-loop experiments and advanced brain-machine interfaces.
• Existing point process filters are computationally intensive for multidimensional problems.
• Accurate decoding of population spiking activity is essential for understanding neural computations.

Purpose of the Study:

• To develop a computationally efficient and accurate approximate filter solution for multidimensional point process problems.
• To enable real-time decoding of neural population activity.
• To extend the application to marked point processes, including both sorted and unsorted spike data.

Main Methods:

• Proposed an approximate filter solution for point process problems using a Gaussian Mixture Model for posterior distributions.
• Developed an efficient algorithm to estimate mixture components, weights, means, and covariances.
• Utilized mixture dropping and merging algorithms to control computational complexity.

Main Results:

• The approximate filter achieved real-time performance, significantly outperforming the exact solution in speed (over 20x in 1D, over 4,000x in 2D).
• Accuracy remained comparable to the exact solution, with minor drops in Root Mean Square Error (RMSE) and highest probability coverage area (HPD).
• Successfully applied the method to decode rat position in 1D and 2D spaces using hippocampal place cell data.

Conclusions:

• The proposed approximate point process filter offers a computationally efficient and accurate solution for real-time neural decoding.
• The methodology is versatile, applicable to both clusterless and sorted spike data.
• This advancement facilitates real-time neural data analysis for neuroscience research and brain-machine interface development.