Reproduce “EEG classification for visual brain decoding with spatio-temporal and transformer based paradigms” on OpenResearch

About this paper

Akanksha Sharma, Jyoti Nigam, Abhishek Rathore, Arnav Bhavsar

Researchers at the Indian Institute of Technology-Mandi introduced hybrid deep learning models for visual brain decoding using EEG, integrating 1D CNNs with either Bi-LSTMs or Transformer encoders. Their CNN-BiLSTM model achieved 71% classification accuracy on the filtered EEG-ImageNet dataset, establishing a new benchmark for this task, while also demonstrating that the learned representations distinguish visual categories.

Key insights

• Hybrid deep learning architectures, combining Convolutional Neural Networks for initial feature extraction with sequential (Bi-LSTM) or attention-based (Transformer) layers, are highly effective for classifying complex spatio-temporal EEG signals.
• The use of a window-based classification approach with majority voting and a rigorously filtered dataset (EEG-ImageNet 2020 release) is crucial for achieving robust and reliable performance in visual brain decoding.
• Learned embeddings from the CNN-BiLSTM model demonstrate clear visual clustering of different image categories, which correlates with distinct brain activity patterns observed in topographic maps, enhancing model interpretability.

Summary by alphaXiv.