Deciphering EEG Waves for the Generation of Images

Research on electroencephalography (EEG) waves has been a crucial topic in the development of BCI technology and neurolinguistics. EEG waves have found been to not only study neuronal activities but also explore human cognitive activities, in diagnoses of neurological disorders and also in the field of criminology to understand underlying psychology. Deciphering these waves to generate images with a focus on open vocabulary stands as an unexplored realm. The pursuit starts with an important step of meticulous pre-processing and analysis of EEG waves using various techniques such as Downsampling, Band Pass Filtering, removal of bad channels, and Independent Component Analysis (ICA). The next target is the EEG-To-Text decoding focusing on open vocabulary to generate images for better communication using the ZuCo 1.0 and ZuCo 2.0 datasets. For text generation tasks, pre-trained large language models such as BERT and BART, and for image generation, a Stable Diffusion model is leveraged. Our model achieves a 0.34 BLEU-1 score on EEG-To-Text decoding. Nevertheless, the model can handle various kinds of data, showing great potential for a high-performance open vocabulary brain-to-text system once an appropriate dataset is available.