Projection-Free CLIP-Scale EEG Latents via a U-Net-Style Autoencoder

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Electroencephalography (EEG) is a key modality for conditioning generative visual models.
• Current representation learning methods use high-capacity autoencoders, leading to issues like representation collapse and loss of signal dynamics, especially under dimensional constraints for vision-language models.

Purpose of the Study:

• To propose a lightweight, projection-free autoencoder for EEG representation learning.
• To generate compact, CLIP-scale latent vectors directly from EEG data.
• To address representation collapse and preserve signal dynamics in EEG-based generative models.

Main Methods:

• Developed a U-Net-style autoencoder architecture.
• Incorporated 1D convolutional residual blocks for temporal dynamics and inter-channel attention for spatial dependencies.
• Utilized skip connections for stable signal reconstruction.
• Trained the model to output latent vectors aligned with the CLIP embedding space.

Main Results:

• The proposed autoencoder successfully tracks complex EEG signal amplitudes without collapsing.
• Achieved superior signal reconstruction fidelity in time and frequency domains under strict dimensional constraints.
• Significantly fewer parameters were required compared to traditional masked autoencoder baselines.
• Preserved partial semantic organization in latent space, outperforming collapsed baseline representations in zero-shot retrieval tasks.

Conclusions:

• The lightweight autoencoder effectively preserves EEG signal structure and dynamics.
• Demonstrates preliminary, above-chance semantic alignment for brain-driven generative pipelines.
• Offers a more efficient and effective approach to EEG representation learning compared to existing methods.