Deep learning-based spike sorting: a survey.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Deep learning is increasingly applied to neuroscience signal processing, particularly for extracellular recordings.
• Spike sorting is essential for assigning action potentials (spikes) to individual neurons from population recordings.

Purpose of the Study:

• To critically synthesize findings from recent deep learning-based spike sorting methodologies.
• To provide an in-depth evaluation of current state-of-the-art approaches, methodologies, and outcomes.

Main Methods:

• Examined 24 articles on deep learning-based spike sorting published up to December 2023.
• Categorized methods into spike detection, feature extraction, and classification, including integrated systems.

Main Results:

• Multi-channel data models show promise, with efficient hardware implementations.
• Convolutional neural networks excel in spike detection and classification due to spatiotemporal processing.
• Autoencoders are used for dimensionality reduction, and integrated systems offer end-to-end solutions.

Conclusions:

• Deep neural networks show significant potential in addressing spike sorting challenges.
• The survey highlights model capabilities and potential biases, serving as a resource for researchers.
• This work aims to inspire future developments in deep learning for neural signal processing.