Comparison of Classifier Architectures for Online Neural Spike Sorting

Area of Science:

• Neuroscience
• Computer Engineering
• Signal Processing

Background:

• High-density intracranial recordings necessitate spike sorting for neuronal spike association.
• On-chip real-time processing for spike sorting faces scalability challenges due to high computational demands of classifiers.

Purpose of the Study:

• To analyze popular classifiers and propose novel hardware architectures for on-chip spike sorting.
• To evaluate proposed architectures based on accuracy and resource requirements for off-chip training and on-chip classification.

Main Methods:

• Review and analysis of several popular classification algorithms.
• Proposal of five new hardware architectures for spike sorting: Support Vector Classification, Fuzzy C-Means, Self-Organizing Maps, Moving-Centroid K-Means, and Cosine Distance Classification.
• Performance evaluation focusing on accuracy and computational resource utilization.

Main Results:

• Self-Organizing Maps (SOM) based neural network classifier identified as the most viable solution.
• A SOM-based spike sorter utilizes only 7.83% of the computational resources compared to hierarchical adaptive means.
• The SOM classifier achieves 3% better accuracy at 7 dB Signal-to-Noise Ratio (SNR).

Conclusions:

• The Self-Organizing Maps classifier presents a highly efficient and accurate approach for on-chip spike sorting.
• Hardware architectures utilizing off-chip training with on-chip classification are crucial for scalable spike sorting.
• The proposed SOM-based architecture significantly reduces computational load while enhancing classification performance.