How much EEG is needed for deep learning with convolutional neural networks? Predicting the benefit from additional...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) classification is crucial for various neurological applications.
• Limited labeled data often constrains the performance of machine learning models in EEG analysis.
• Understanding data requirements is essential for efficient EEG study design.

Purpose of the Study:

• To quantify the impact of training data size on EEG classification accuracy using convolutional neural networks.
• To evaluate parametric models for extrapolating neural network performance to larger EEG datasets.
• To provide insights into data acquisition strategies for EEG research.

Main Methods:

• Evaluated three neural network architectures across three EEG classification tasks.
• Systematically varied the number of subjects and EEG data duration per subject.
• Assessed eight parametric models for fitting and extrapolating learning curves, analyzing prediction error and uncertainty.

Main Results:

• Learning curve characteristics (slope, shape, asymptotic performance) varied significantly across tasks but were consistent across network architectures.
• Reliable extrapolation of performance using scaling laws necessitated data from several hundred subjects.
• The benefit of increasing EEG recording length per subject reached a plateau after only a few seconds.

Conclusions:

• Task-specific learning curves can guide EEG study design and data acquisition.
• Extrapolated learning curves aid in cost-benefit analyses for acquiring labeled EEG data.
• Findings highlight the substantial data requirements for robust EEG classification and the diminishing returns of extended recording durations.