Reproduce “DEAP DIVE: Dataset Investigation with Vision transformers for EEG evaluation” on OpenResearch

About this paper

Annemarie Hoffsommer, Helen Schneider, Svetlana Pavlitska, J. Marius Zöllner

A framework employing Continuous Wavelet Transformation (CWT) and Vision Transformers (ViT) demonstrates the capability for accurate human emotion prediction from electroencephalogram (EEG) data using fewer channels. The approach achieved 91.57% classification accuracy with 12 Emotiv channels and set a baseline for continuous emotion regression with an RMSE of 0.57 using 32 EEG channels on the DEAP dataset.

Key insights

• Continuous Wavelet Transformation (CWT) effectively transforms raw EEG signals into visual scaleograms, enabling Vision Transformers (ViT) to capture complex temporal and frequency features for emotion recognition.
• High emotion classification accuracy (91.57%) is achievable with a substantially reduced number of EEG channels (12 channels), demonstrating the practical feasibility of low-cost, portable EEG devices.
• The study revealed that increasing the number of input channels does not guarantee improved performance, with optimized subsets sometimes outperforming configurations using all available channels.

Summary by alphaXiv.