Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding

Baimam Boukar Jean Jacques1, Brandone Fonya1, Nchofon Tagha Ghogomu1, Pauline Nyaboe1
1Carnegie Mellon University
Paper arXiv Video Code Data

EEG Neural Decoding

Cross-Subject & Cross-Task Generalization

Zero-Shot Neural Priors enable robust EEG decoding across subjects and tasks without fine-tuning.

Key Contributions

🧠 Zero-Shot Learning

No subject-specific training required for new participants

🔄 Cross-Task Transfer

Generalizes across different cognitive paradigms

âš¡ Neural Priors

Learned universal representations from diverse datasets

Abstract

The development of generalizable electroencephalography (EEG) decoding models is a crucial step toward robust brain-computer interfaces and objective neural biomarkers for mental health. EEG decoding has historically been constrained by the lack of scalable models that generalize across individuals and cognitive tasks.

Our project is motivated by the desire to address this gap by developing a zero-shot learning approach for cross-subject EEG decoding. We aim to build a large pretrained model that captures domain- and subject-invariant neural representations to enable behavioral prediction, leveraging the large-scale Healthy Brain Network (HBN) dataset.

Our work contributes to the growing effort to establish EEG as a robust tool for objective neuroscience and computational psychiatry, with potential applications in the development of EEG-based biomarkers for mental health, clinical assessment and adaptive brain-computer interfaces. The project repository can be accessed at: https://github.com/Nchofon/neuronium

Method Overview

Method Visualization

Architecture diagram and results will be displayed here

Cross-Subject Generalization

Our method achieves robust performance across different subjects without requiring any subject-specific training or calibration, demonstrating the effectiveness of neural priors.

Cross-Subject Performance Chart

Cross-Task Transfer

The learned neural priors transfer effectively across different cognitive tasks, enabling zero-shot decoding for novel experimental paradigms.

Cross-Task Results

Experimental Results

Performance Comparison

We evaluate our method on multiple benchmark EEG datasets and compare against state-of-the-art approaches. Our zero-shot neural priors consistently outperform existing methods in cross-subject scenarios.

Performance Comparison Table

Accuracy Comparison Chart


Ablation Studies

We conduct comprehensive ablation studies to analyze the contribution of each component in our neural architecture, demonstrating the importance of the proposed neural priors.

Ablation Study Results

Related Work

Our work builds upon recent advances in EEG signal processing and cross-domain transfer learning.

Recent work on Domain-Adaptive EEG Decoding has shown promising results for subject adaptation, but requires target domain data for fine-tuning.

Deep Transfer Learning for BCI and Neural Domain Adaptation in Brain Signals explore similar ideas but focus on supervised adaptation approaches.

Some works address cross-subject variability through Riemannian Geometry approaches, Adversarial Domain Adaptation, and Meta-Learning for EEG.

For a comprehensive overview of current EEG decoding methods, see Recent Advances in EEG Signal Processing and Cross-Subject BCI: A Survey.

BibTeX

@article{zeroshoteeg2025,
  author    = {Baimam Boukar Jean Jacques and Brandone Fonya and Nchofon Tagha Ghogomu and Pauline Nyaboe},
  title     = {Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding},
  institution = {Carnegie Mellon University},
  year      = {2025},
}