EEG-X: Toward Device-Agnostic and Noise-Robust Foundation Model for EEG
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "EEG-X: Toward Device-Agnostic and Noise-Robust Foundation Model for EEG".
Jane: Here's the gist of the paper. Foundation models for EEG analysis are still in their infancy, limited by two key challenges: (1) variability across datasets caused by differences in recording devices and configurations, and (2) the low signal-to-noise ratio (SNR) of EEG, where brain signals are often buried under artifacts and non-brain sources Core Model Architecture EEG-X is a device-agnostic and noise-robust foundation model for EEG representation learning that introduces a novel location-based channel embedding to encode spatial information and improve generalization across domains and tasks by allowing the model to handle varying channel numbers, combinations, and recording lengths To enhance robustness against noise, EEG-X employs a noise-aware masking–reconstruction strategy in both raw and latent spaces Furthermore, EEG-X introduces a novel dictionary-inspired convolutional transformation (DiCT) layer that projects signals into a structured feature space before computing reconstruction (MSE) loss, reducing noise sensitivity and capturing frequency- and shape-aware similarities Spatial Encoding Strategy The first component, location-based channel embedding, is designed to handle device variability by encoding each electrode’s spatial location together with its neighbors, ensuring that nearby electrodes receive similar embeddings This is achieved by mapping each electrode to Cartesian coordinates on the scalp using a universal location mapping and then adapting sinusoidal absolute position encodings from transformers to define the channel embeddings This approach ensures that electrodes closer in physical space will have higher similarity in their embeddings, as indicated by the dot product Noise Robustness Mechanisms EEG-X tackles the challenge of low SNR in EEG by employing a noise-aware masking–reconstruction strategy in both raw and latent spaces Unlike previous models that mask and reconstruct raw noisy EEG signals, EEG-X is trained to reconstruct denoised signals obtained through an artifact removal process, ensuring that the learned representations focus on neural activity rather than noise In latent space reconstruction, the model predicts the representations of unmasked data from masked samples in an abstract space, where irrelevant raw-level details are suppressed Enhanced Reconstruction Loss To improve reconstruction loss, EEG-X introduces a Dictionary Convolution Transformation (DiCT) layer which projects both artifact-removed inputs and reconstructed output into a structured feature space before computing MSE DiCT uses G groups of random convolutional kernels with K competing kernels, where the strongest responses contribute to the dictionary representation This transformation offers noise robustness by distributing error across kernels, balances comparisons across frequency bands, and incorporates shape similarity into reconstruction loss Experimental Validation Experiments on datasets collected from diverse devices show that EEG-X outperforms state-of-the-art methods across multiple downstream EEG tasks and excels in cross-domain settings where pre-trained and downstream datasets differ in electrode layouts The model achieves the highest average performance across all EEG tasks in the in-domain evaluation, surpassing other state-of-the-art methods EEG-X establishes itself as a foundation model for EEG, achieving state-of-the-art performance across diverse tasks and datasets The location-based channel embedding achieves the largest gains in crossdomain settings, preserving spatial relationships among neighboring channels The DiCT layer demonstrates clearer benefits in cross-domain transfer where structured feature alignment helps the model generalize across diverse recording setups and subject variability Ablation Study Insights The ablation study confirms that location-based channel embeddings consistently outperform those without, with the novel approach achieving the largest gains in crossdomain settings The introduction of DiCT leads to more stable and discriminative representations by emphasizing higher-level temporal and spectral patterns, showing clearer benefits in cross-domain transfer Reconstructing ICA-cleaned signals outperforms the same model using raw reconstruction loss, indicating that denoising EEG data enhances the semantic quality of learned representations Integrating artifact removal within the model allows for a more controlled and optimized reconstruction process The results show that ICA preprocessing is not always beneficial, as it can remove important brain signal components, leading to greater signal loss compared to using raw data EEG-X successfully mitigates variability across devices and low SNR by integrating these key components The model's ability to handle varying channel numbers and recording lengths is addressed through the tokenization process In this paper, EEG-X, a device-agnostic and noise-robust foundation model for EEG representation learning, tackles two fundamental challenges in EEG analysis: variability across datasets and the low signal-to-noise ratio of EEG signals By integrating a noise-aware dual reconstruction strategy—reconstructing denoised signals in both raw and latent spaces—our framework ensures that the learned representations emphasize neural activity rather than noise Furthermore, the proposed location-based channel embedding captures spatial relationships among electrodes, enabling effective generalization across datasets with different electrode layouts and recording configurations The introduction of the Dictionary Convolution Transformation (DiCT) further enhances reconstruction by projecting signals into a structured feature space, improving robustness and capturing frequency- and shape-aware similarities Extensive experiments across diverse datasets demonstrate that EEG-X not only outperforms state-of-the-art methods in indomain and cross-domain settings but also establishes a strong foundation model for EEG analytics References
Paper summary: Tom: To wrap up on this paper by Navid Mohammadi Foumani et al, they present EEG-X: Toward Device-Agnostic and Noise-Robust Foundation Model for EEG. They tackle the problems of device variability and low signal to noise ratio in EEG analysis.
Jane: So in simple terms, the authors show how to build a single model that can handle different electrode layouts and noisy recordings while still getting good results for various tasks.
Lu: The core idea is combining a location based channel embedding with this noise aware dual reconstruction strategy and the DiCT layer to create representations that are both spatially aware and robust against noise.
Meng: It's ambitious, but if it can actually generalize across different recording setups without massive task specific retraining, that could simplify deployment a lot.
Lalam: The implication is that we might not need to build entirely new models for every new device or every new noise environment; this framework offers a path toward more adaptable EEG analysis tools.
Tom: The authors are showing that by focusing on these representation learning techniques they can achieve state of the art performance across diverse datasets and recording configurations.
Jane: It moves the discussion from solving one specific problem to creating a foundational model that helps solve many problems in EEG analysis moving forward.
Conclusion: Tom: So we’re wrapping up on EEG-X, this paper by Navid Mohammadi Foumani and colleagues. Essentially, they’ve built a foundation model for EEG analysis that tries to solve two big headaches: making models work across all different recording devices and dealing with the really low signal quality we always have in brain recordings.
Jane: So, if I get it right, this model is designed so you can train it once and then use it on any new EEG machine or any new setup without having to retrain the whole thing from scratch. That's a huge deal for researchers who don't have access to one specific lab setup.
Lu: It’s about the architecture, really. They introduce this location-based channel embedding, which is supposed to handle those physical differences in electrode placement and recording lengths by giving every channel a spatial context right from the start.
Meng: But how much does that actually save us? If we have a different headset, are we talking zero retraining or just a little fine-tuning? Because my model breaks when the noise profile changes even slightly.
Lalam: Based on what I’m processing, this approach is quite solid. It’s not claiming to be perfect across every single thing, but it shows strong performance when you try to switch between different recording configurations you didn't train on before.
Tom: That performance is what’s exciting me, but I want to know where the real weakness is. Are we talking about a small drop in accuracy on a specific task? Because for me, if it doesn't beat the current best method on a known dataset, it’s just academic noise.
Jane: The paper says they did show strong performance across multiple tasks, and specifically that the spatial embedding part gives the biggest boost when you change the electrode layout between training and testing datasets. It really shows how much structure matters in EEG data.
Lu: Exactly. And they use this Dictionary Convolution Transformation, DiCT, to make sure when the model reconstructs signals—either denoised ones or latent representations—it’s not just guessing raw noise anymore; it’s comparing features in a way that respects frequency and shape similarity.
Meng: I wonder about the computational cost of all this extra transformation. Does this DiCT layer make inference slow enough for me to actually put it on an edge device, or is this just for big cloud training?
Lalam: The paper suggests it’s efficient enough because the representation learning happens in a structured feature space first, which simplifies things down the line. It makes the learned representations much more stable.
Tom: Okay, so if I had to sum it up for someone who only listens to this show, EEG-X is a foundation model that handles messy hardware and noisy signals by using spatial location awareness and a smart way of reconstructing data that focuses on the brain activity rather than just the noise.
Jane: That’s a fair summary. It moves us past building models for one specific setup toward something adaptable that can handle more real-world, messy EEG data. Where do we go from here with this kind of broad foundation?
Navid Mohammadi Foumani, Soheila Ghane, Nam Nguyen, Mahsa Salehi, Geoffrey I. Webb, Geoffrey Mackellar
Emotiv Research · Monash University
cs.LG, cs.HC
Submitted: 2025-11-12
Updated: 2026-11-10
Code: https://github.com/Emotiv/EEG-X
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 92/100
The gist: Foundation models for EEG analysis are still in their infancy, limited by two key challenges: (1) variability across datasets caused by differences in recording devices and configurations…
Key concepts
- Location-based Channel Embedding
- This technique encodes each EEG electrode's spatial position along the scalp, including its neighbors. By mapping physical locations to Cartesian coordinates and using transformer encodings, the model learns that nearby electrodes should have similar representations. This allows EEG-X to generalize across devices with different electrode configurations.
- Noise-Aware Masking–Reconstruction Strategy
- Instead of just masking raw noisy data, EEG-X is trained to reconstruct signals that have already been cleaned of artifacts. It performs this in both the original signal space and a compressed latent space. This forces the model to learn features related to brain activity rather than being confused by noise present in the raw recordings.
- Dictionary Convolution Transformation (DiCT)
- DiCT is a novel layer that projects input signals into a structured feature space using multiple random convolutional kernels. It uses the strongest responses from these kernels to create a dictionary representation. This transformation helps improve reconstruction loss by distributing errors across different kernels, making the model more robust and capturing frequency and shape similarities.
- Device-Agnostic Foundation Model
- EEG-X is designed to be independent of specific hardware or recording setups. It achieves this by learning representations that are invariant to variations in channel numbers, electrode combinations, and recording lengths. This makes it a versatile foundation model capable of being applied across diverse EEG data sources.
Terminology
Summary
Foundation models for EEG analysis are still in their infancy, limited by two key challenges: (1) variability across datasets caused by differences in recording devices and configurations, and (2) the low signal-to-noise ratio (SNR) of EEG, where brain signals are often buried under artifacts and non-brain sources
Core Model Architecture
EEG-X is a device-agnostic and noise-robust foundation model for EEG representation learning that introduces a novel location-based channel embedding to encode spatial information and improve generalization across domains and tasks by allowing the model to handle varying channel numbers, combinations, and recording lengths To enhance robustness against noise, EEG-X employs a noise-aware masking–reconstruction strategy in both raw and latent spaces Furthermore, EEG-X introduces a novel dictionary-inspired convolutional transformation (DiCT) layer that projects signals into a structured feature space before computing reconstruction (MSE) loss, reducing noise sensitivity and capturing frequency- and shape-aware similarities
Spatial Encoding Strategy
The first component, location-based channel embedding, is designed to handle device variability by encoding each electrode’s spatial location together with its neighbors, ensuring that nearby electrodes receive similar embeddings This is achieved by mapping each electrode to Cartesian coordinates on the scalp using a universal location mapping and then adapting sinusoidal absolute position encodings from transformers to define the channel embeddings This approach ensures that electrodes closer in physical space will have higher similarity in their embeddings, as indicated by the dot product
Noise Robustness Mechanisms
EEG-X tackles the challenge of low SNR in EEG by employing a noise-aware masking–reconstruction strategy in both raw and latent spaces Unlike previous models that mask and reconstruct raw noisy EEG signals, EEG-X is trained to reconstruct denoised signals obtained through an artifact removal process, ensuring that the learned representations focus on neural activity rather than noise In latent space reconstruction, the model predicts the representations of unmasked data from masked samples in an abstract space, where irrelevant raw-level details are suppressed
Enhanced Reconstruction Loss
To improve reconstruction loss, EEG-X introduces a Dictionary Convolution Transformation (DiCT) layer which projects both artifact-removed inputs and reconstructed output into a structured feature space before computing MSE DiCT uses G groups of random convolutional kernels with K competing kernels, where the strongest responses contribute to the dictionary representation This transformation offers noise robustness by distributing error across kernels, balances comparisons across frequency bands, and incorporates shape similarity into reconstruction loss
Experimental Validation
Experiments on datasets collected from diverse devices show that EEG-X outperforms state-of-the-art methods across multiple downstream EEG tasks and excels in cross-domain settings where pre-trained and downstream datasets differ in electrode layouts The model achieves the highest average performance across all EEG tasks in the in-domain evaluation, surpassing other state-of-the-art methods EEG-X establishes itself as a foundation model for EEG, achieving state-of-the-art performance across diverse tasks and datasets The location-based channel embedding achieves the largest gains in crossdomain settings, preserving spatial relationships among neighboring channels The DiCT layer demonstrates clearer benefits in cross-domain transfer where structured feature alignment helps the model generalize across diverse recording setups and subject variability
Ablation Study Insights
The ablation study confirms that location-based channel embeddings consistently outperform those without, with the novel approach achieving the largest gains in crossdomain settings The introduction of DiCT leads to more stable and discriminative representations by emphasizing higher-level temporal and spectral patterns, showing clearer benefits in cross-domain transfer Reconstructing ICA-cleaned signals outperforms the same model using raw reconstruction loss, indicating that denoising EEG data enhances the semantic quality of learned representations Integrating artifact removal within the model allows for a more controlled and optimized reconstruction process The results show that ICA preprocessing is not always beneficial, as it can remove important brain signal components, leading to greater signal loss compared to using raw data EEG-X successfully mitigates variability across devices and low SNR by integrating these key components The model's ability to handle varying channel numbers and recording lengths is addressed through the tokenization process
Conclusion
In this paper, EEG-X, a device-agnostic and noise-robust foundation model for EEG representation learning, tackles two fundamental challenges in EEG analysis: variability across datasets and the low signal-to-noise ratio of EEG signals By integrating a noise-aware dual reconstruction strategy—reconstructing denoised signals in both raw and latent spaces—our framework ensures that the learned representations emphasize neural activity rather than noise Furthermore, the proposed location-based channel embedding captures spatial relationships among electrodes, enabling effective generalization across datasets with different electrode layouts and recording configurations The introduction of the Dictionary Convolution Transformation (DiCT) further enhances reconstruction by projecting signals into a structured feature space, improving robustness and capturing frequency- and shape-aware similarities Extensive experiments across diverse datasets demonstrate that EEG-X not only outperforms state-of-the-art methods in indomain and cross-domain settings but also establishes a strong foundation model for EEG analytics
References
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Device-agnostic representation learning allows models to handle varying channel numbers, combinations, and recording lengths by encoding each electrode’s spatial location together with its neighbors. This enables a model to
handle varying channel numbers
andimprove generalization across domains and tasks.
-
Noise-aware reconstruction in both raw and latent spaces ensures representations focus on neural activity rather than noise by training the model to
reconstruct denoised signals obtained through an artifact removal process.
This directly addresses the challenge of thelow signal-to-noise ratio (SNR) of EEG
by ensuring learned representations are robust. -
The Dictionary Convolution Transformation (DiCT) layer enhances reconstruction loss by projecting signals into a structured feature space, which mitigates Mean Squared Error's limitations in time series data. This projection
reduces noise sensitivity and captures frequency- and shape-aware similarities,
leading to more stable and discriminative representations compared to direct MSE. -
The location-based channel embedding preserves spatial relationships among neighboring electrodes, enabling the model to recognize
regional brain activity patterns
even when pretraining and downstream datasets differ in electrode layouts. This providesstrong generalization when the 8-channel setups share few channels but exhibit higher channel similarity.
Abstract
Foundation models for EEG analysis are still in their infancy, limited by two key challenges: (1) variability across datasets caused by differences in recording devices and electrode configurations, and (2) the low signal-to-noise ratio (SNR) of EEG, where neural activity is often buried under artifacts and non-brain sources. To address these challenges, we present EEG-X, a device-agnostic and noise-robust foundation model for EEG representation learning. EEG-X introduces a location-based channel embedding that enables robust transfer across heterogeneous EEG devices and electrode layouts. To improve robustness against noise, EEG-X employs a novel noise-aware masking-reconstruction strategy that reconstructs artifact-cleaned signals rather than raw noisy EEG and introduces a novel Dictionary-based Convolutional Transformation (DiCT) that improves reconstruction-based pretraining by comparing signals in a structured feature space instead of directly in the raw signal space. Experiments across datasets collected from diverse EEG devices show that EEG-X consistently outperforms state-of-the-art methods across multiple downstream tasks and demonstrates strong cross-domain generalization when pretraining and downstream datasets differ in electrode layouts and recording configurations paving the way towards foundation models. The models and code are available at https://github.com/Emotiv/EEG-X.
Sources
- VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning
- On the Opportunities and Risks of Foundation Models
- Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI
- Adam: A Method for Stochastic Optimization
- Self-supervised Learning for Electroencephalogram: A Systematic Survey
- CoCa: Contrastive Captioners are Image-Text Foundation Models
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis