Decoding visual brain representations from electroencephalography through knowledge distillation and latent diffusion...

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) are advancing rapidly, with a focus on decoding visual information from brain activity.
• Electroencephalography (EEG) offers a non-invasive method for capturing neural signals, but decoding complex visual data remains challenging.

Purpose of the Study:

• To develop an innovative method for training an EEG classifier using knowledge distillation.
• To reconstruct images from EEG data, enabling direct brain decoding of visual stimuli.
• To evaluate the performance of the proposed method against existing benchmarks.

Main Methods:

• EEG data from participants viewing images from ImageNet and THINGS-EEG 2 datasets were analyzed.
• EEG signals were converted into spectrograms and used to train a convolutional neural network (CNN).
• Knowledge distillation from a pre-trained Contrastive Language-Image Pre-Training (CLIP)-based teacher network was employed, alongside latent diffusion models for image reconstruction.

Main Results:

• The EEG classifier achieved a top-5 accuracy of 87%, significantly outperforming standard CNNs and recurrent neural network (RNN) benchmarks.
• The method successfully reconstructed images directly from EEG data.
• The approach demonstrated superior performance in brain decoding tasks.

Conclusions:

• The proposed knowledge distillation and latent diffusion model architecture enables accurate image decoding and reconstruction from EEG.
• This brain decoding technique has significant implications for developing advanced BCIs and personalized feedback systems.
• The study paves the way for rapid, individualized feedback experiments in BCI research.