Skip to content

Learning Robust Deep Visual Representations from EEG Brain Recordings

🔥INFO

Blog: 2025/07/28 by IgniSavium

  • Title: Learning Robust Deep Visual Representations from EEG Brain Recordings
  • Authors: Prajwal Singh, Shanmuganathan Raman (IIT Gandhinagar, India)
  • Published: October 2023
  • Comment: WACV
  • URL: https://openaccess.thecvf.com/content/WACV2024/papers/Singh_Learning_Robust_Deep_Visual_Representations_From_EEG_Brain_Recordings_WACV_2024_paper.pdf

🥜TLDR: Encoder Training: additional CLIP distillation

Motivation

This paper aims to overcome the limitations of low-quality image synthesis and heavy reliance on label supervision in EEG-based image generation by proposing two-stage framework (EEGStyleGAN-ADA) that significantly improves synthesis quality and generalizability across datasets compared to prior state-of-the-art methods.

Model

Architecture

image-20250729173333788

  1. Train the EEG feature encoder by triplets loss.

image-20250729172425488

  1. Fine-tune the EEG feature encoder by standard CLIP loss.

image-20250729173432920

  1. Use the CLIP-aligned EEG feature as StyleGAN-ADA input.

Evaluation

Pre-training Effectiveness

triplet loss vs. supervised classification loss linear separability

image-20250729174606924

image-20250729174622100

image-20250729175259719

image-20250729174043802

Image Synthesis Performance

image-20250729175654015

Data simulation

image-20250729180232550

Image Retrieval Validation

image-20250729180429112