Deep Metric Learning With Locality Sensitive Mining for Self-Correcting Source Separation of Neural Spiking Signals.

Area of Science:

• Neuroengineering
• Computational Neuroscience
• Machine Learning

Background:

• Automated source separation algorithms are crucial for decomposing neurophysiological signals into spiking sources.
• Noisy or complex recordings lead to errors in these algorithms, degrading performance and requiring manual correction.

Purpose of the Study:

• To propose an automated error correction methodology for neurophysiological signal decomposition.
• To investigate deep metric learning (DML) techniques for identifying and assigning spiking events.
• To develop an improved DML approach for accurate error correction.

Main Methods:

• Utilized a deep metric learning (DML) framework to create embedding spaces for spiking events.
• Investigated different DML techniques for preserving intraclass semantic structure.
• Introduced locality sensitive mining, a sampling-based augmentation for DML losses.

Main Results:

• The proposed DML framework successfully generates embedding spaces for identifying and assigning spiking events.
• Locality sensitive mining significantly enhances the local semantic structure of the embedding space.
• The method automatically identifies incorrectly labeled spiking events with high accuracy.

Conclusions:

• The developed automated error correction methodology using DML and locality sensitive mining is effective.
• This approach reduces the need for manual cleaning of neurophysiological data.
• The findings advance automated analysis in neuroengineering and neuroscience.