Trainee Project
Title:
Deep learning for action potential features extraction
Dates:
2025/03/03 - 2025/09/19
Student:
Supervisor(s):
Description:
Action potentials (or spikes) form the foundation of neuronal communication, enabling the transmission of information within brain networks. These electrical activities, generated by individual neurons, can be recorded using microscopic electrodes implanted within the brain tissue. Each electrode captures signals emitted by multiple neurons. Spike sorting aims to classify spikes in order to isolate and identify the activity of individual neurons. Traditional approaches rely on a thresholding phase to detect events for sorting, followed by feature extraction and the application of classification methods.
The use of deep learning models to address this problem has been minimally explored [1,2], mainly due to the variability of action potentials and the need for a sufficiently large labeled dataset. This project aims to restrict the use of deep learning to the feature extraction phase, where the network will be specifically trained to extract the most discriminative features for classification. These features will then be used as inputs for supervised (on simulated data) and unsupervised classification algorithms. Particular attention will be given to selecting the most suitable network model (network structure) for this issue, as well as to defining an appropriate loss function to identify features general enough to be applicable to new data. Initially, the study will focus on simulated signals in a realistic framework, both in terms of action potential shapes and background activity [3]. The model will then be applied to real signals recorded in humans.
[1] Lee, Jin Hyung, et al. "YASS: yet another spike sorter." Advances in neural information processing systems 30 (2017).
[2] Rácz, Melinda, et al. "Spike detection and sorting with deep learning." Journal of neural engineering 17.1 (2020).
[3] Tran, H., Signatures extracellulaires des potentiels d'action neuronaux: modélisation et analyse, Thèse de l'Université de Lorraine (2019).
The use of deep learning models to address this problem has been minimally explored [1,2], mainly due to the variability of action potentials and the need for a sufficiently large labeled dataset. This project aims to restrict the use of deep learning to the feature extraction phase, where the network will be specifically trained to extract the most discriminative features for classification. These features will then be used as inputs for supervised (on simulated data) and unsupervised classification algorithms. Particular attention will be given to selecting the most suitable network model (network structure) for this issue, as well as to defining an appropriate loss function to identify features general enough to be applicable to new data. Initially, the study will focus on simulated signals in a realistic framework, both in terms of action potential shapes and background activity [3]. The model will then be applied to real signals recorded in humans.
[1] Lee, Jin Hyung, et al. "YASS: yet another spike sorter." Advances in neural information processing systems 30 (2017).
[2] Rácz, Melinda, et al. "Spike detection and sorting with deep learning." Journal of neural engineering 17.1 (2020).
[3] Tran, H., Signatures extracellulaires des potentiels d'action neuronaux: modélisation et analyse, Thèse de l'Université de Lorraine (2019).
Keywords:
Spike sorting, deep learning
Department(s):
| Biology, Signals and Systems in Cancer and Neuroscience |