HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning

arXiv:2610.16589 · cs.LG · Submitted 2026-10-12 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning".

Jane: The gist The proposed HyBRA framework maps heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're starting with the paper "HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning". It basically proposes a framework that maps messy EEG channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic distances.

Jane: That sounds complicated. If I were driving or cooking, I'd wonder what this means for me—does it mean my brain activity looks the same whether I use a different set of electrodes or just change my setup?

Lu: The core idea is that they take those heterogeneous channel observations and convert them into something unified. They use coordinate-aware patch encoding to add the three dimensional electrode coordinate to every single channel patch before processing it through a shared temporal encoder.

Meng: So, if you're thinking about deployment, having a fixed regional token space means once you do this mapping, your downstream models only have to deal with K regions instead of potentially hundreds of channels that change every time you switch recording setups.

Lalam: The self-supervised pretraining part involves masking tokens from Equation nine and using the Region-Time Transformer to predict those masked latent features, evaluated by a cosine loss between the prediction and the original pre-mask token.

Tom: And they add a raw signal reconstruction loss based on a channel-to-region prior to make sure those region tokens are actually capturing meaningful signal information from the input patches.

Jane: So, it’s not just about getting a good representation; it’s about building this self-supervised pretraining objective that makes those regional tokens actually useful for prediction tasks.

The paper's summary: Tom: Now let's talk about what they actually did in "HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning". They point out several specific ways this method improves things.

Jane: They show that removing the coordinate encoding lowers the balanced accuracy by one point six seven percentage points on TUAB and five point eight five points on TUEV, which suggests the spatial information is important but maybe not everything can be perfectly encoded beforehand.

Lu: And they also test different matching functions; they show that replacing hyperbolic matching with Euclidean or cosine variants lowers the mean balanced accuracy on both datasets, which supports their choice of a feature-dependent hyperbolic distance.

Meng: Crucially, removing the channel-evidence injection branch reduces the balanced accuracy by one point zero one and six point seven two points, which really supports their claim that this complementary evidence pathway is necessary to preserve information from the original channel tokens during aggregation.

Lalam: They also find that the complete Region-Time Transformer performs best because modeling temporal dynamics within each region and the interaction across regions provides complementary information.

Tom: The results are pretty compelling. On OpenBMI-MI, HyBRA gets a comparable balanced accuracy to CSBrain at eighty-one point one seven versus eighty-one point zero four, but it also improves the AUROC and AUPRC by one point eight four and five point two three percentage points on that specific dataset.

Jane: It suggests that for EEG representation learning, you need to respect both where you are physically located and what features are actually important within your data.

The paper's improvements: Tom: So wrapping up the HyBRA-EEG paper, the main point is that combining a montage-specific region prior with feature-dependent hyperbolic distance lets you get the highest mean balanced accuracy across six benchmark datasets.

Jane: And they also showed that including that channel evidence injection branch is important for preserving finegrained discriminative cues during aggregation, and using Euclidean or cosine variants doesn't beat the hyperbolic matching in some cases.

Lu: It suggests that this approach of mapping heterogeneous channel observations into a fixed regional token space is effective because it successfully manages the variability introduced by different electrode montages and recording settings.

Meng: From a production angle, if you can get these fixed regional tokens, it simplifies the downstream architecture significantly while retaining most of the predictive power from the raw data.

Lalam: The paper concludes that HyBRA achieves this by combining short-term channel features with electrode coordinates and mapping them into a fixed region-token space using the Hyperbolic Region Mixer >.

Tom: It's a solid framework, especially when you look at how it outperforms baselines like LaBraM on Mental Arithmetic while using only one quarter of the parameters.

Jane: It points toward future work looking at adaptive region construction and seeing if this works well across different centers or electrode configurations.

Conclusion: Tom: So HyBRA is basically taking all those messy EEG channel observations and mapping them onto a fixed regional token space using electrode coordinates, region priors, and hyperbolic distances to learn better self-supervised representations.

Jane: It’s about taking data that changes depending on where you put the electrodes—the montage—and making it consistent so the AI can actually learn the brain patterns instead of just learning the specific setup.

Lu: The core idea is this coordinate-aware patch encoding, where you add the three dimensional electrode coordinate to every single channel patch before passing it to a shared temporal encoder.

Meng: I’m interested in how they handle that mapping because if we can fix that regional space, maybe deploying these models across different hospitals won't require retraining from scratch every time.

Lalam: The Hyperbolic Region Mixer is where the magic happens; it maps those channel tokens and learnable region prototypes onto the Poincaré ball using feature-dependent hyperbolic distance, which calculates matching weights based on a softmax function involving alpha, beta, c, k, and p.

Jane: That weighting mechanism with alpha beta c k p—it sounds like they’re not just doing a simple average; they’re letting the features decide how much each channel token matters for a specific region.

Tom: Yeah that feature-dependent distance is what I like; it means the model isn't blindly pooling everything together, which is usually where things get messy.

Lu: And then you get these region tokens, and they feed into a Region-Time Transformer which models temporal dynamics within those regions and also how different regions interact with each other.

Meng: If this works out in production, the cost of running inference should drop because you’re dealing with a fixed set of K regions instead of millions of raw channel features constantly.

Lalam: The self-supervised pretraining uses a binary mask over those position-aware tokens, and they train the Region-Time Transformer to predict those masked tokens using a cosine loss between the prediction and the original pre-mask token.

Jane: And they add that raw signal reconstruction loss based on a channel-to-region prior to make sure those region tokens are actually capturing meaningful signal information from the input patches.

Tom: The experimental results show that HyBRA gets the highest mean balanced accuracy across six downstream datasets, with a macro average of zero point seven two seven eight.

Lu: And they showed that removing coordinate encoding lowered accuracy by one point six seven percentage points on TUAB and five point eight five points on TUEV, which supports the necessity of that spatial embedding to get good results there.

Jane: That’s a big win for showing how much spatial context matters in EEG representation learning, especially when you look at the trade-offs they made with Euclidean or cosine matching variants.

Xin Liu, Feng Lu, Wei Li

cs.LG

Submitted: 2026-10-12

Updated: 2026-10-12

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 85/100

The gist: The gist The proposed HyBRA framework maps heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic

Key concepts

Coordinate-Aware Patch Encoding: This step takes raw EEG signals from each channel, breaks them into small time segments (patches), and adds the 3D physical location of the electrode to each segment. This creates 'coordinate-aware' tokens that embed spatial information directly into the initial data representation.
Hyperbolic Region Mixer: This module takes the coordinate-aware channel tokens and maps them into a fixed set of regional prototypes using hyperbolic distance. It calculates matching weights based on feature-dependent distances, effectively organizing observations into a structured 'region-token space' that supports inter-region modeling.
Channel Evidence Injection: This pathway adds local context to each region token by injecting weighted evidence derived from the original channel features. This ensures that the regional representations retain crucial, fine-grained information from the individual channels during self-supervised training.
Region-Time Transformer: This final component processes the structured region tokens to model both temporal dynamics within each specific brain region and how those different regions interact with each other over time. It uses attention mechanisms to capture these complex spatial and temporal relationships.

Terminology

Aggregation — 1× in this paper
Representation Learning — 1× in this paper · explained in 3 episodes

Summary

The gist The proposed HyBRA framework maps heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic distances to achieve superior self-supervised EEG representation learning.

HyBRA Architecture

Coordinate-Aware Patch Encoding involves dividing each channel into temporal patches and processing them through a shared temporal encoder, where the three-dimensional electrode coordinate is added to every patch from channel c (Page 3).
The resulting coordinate-aware channel tokens He are passed to the Hyperbolic Region Mixer, which maps montage-dependent observations into a fixed region-token space R over a fixed set of K scalp-electrode regions (Page 3).
In the Hyperbolic Region Mixer, the channel tokens and learnable region prototypes are mapped to the Poincaré ball using feature-dependent hyperbolic distance, where matching weights are calculated as alpha b c k p = softmax c∈Ck lc k − λdb c k p (Page 3).
The resulting region-token space forms the global component of HyBRA, which is then processed by a Region-Time Transformer to model temporal dynamics within regions and interactions among regions (Page 3).

Channel Evidence Injection

The complementary channel-evidence pathway forms the local component by drawing on the pre-aggregation channel features and injecting weighted channel-level evidence into each region token (Page 3).

A normalized region-specific linear scorer gives sb c k p = w⊤ k LN heb c p (Page 3).
The evidence weights and region evidence are calculated as η b c k p = softmax c∈Ck sb c k p + lc k e ch b k p (Page 3).
This evidence is injected through a region-specific, feature-wise gate Rb k p = r agg b k p + σ (γk) ⊙ Dropout Wo LN e ch b k p + bo (Page 3).

Self-Supervised Pretraining

HyBRA performs masked region-time modeling by sampling a binary mask M over the position-aware tokens Z0 from Eq. 9, replacing each masked token with a learnable token m (Page 4).

The Region-Time Transformer processes the corrupted region-time grid to yield ZeL = fRT(Ze0) (Page 4).
Predictions are evaluated only at masked positions using the cosine loss between the predicted feature and the pre-mask token at that position, Lfeat (Page 4).
The reconstruction loss Lraw is calculated for a normalized region-level signal derived from the input patches using the channel-to-region prior, and the pretraining objective is LSSL = Lfeat + λrawLraw (Page 4).

Experimental Results

HyBRA achieves the highest mean BAcc on all six downstream datasets, with a macro-averaged BAcc of 0.7278 (Page 2).

Against the strongest pretrained baseline on each dataset, the BAcc gains are reported as 2.37, 1.79, and 6.99 percentage points on TUEV, BCIC-IV 2a, and MentalArithmetic (Page 5).
On OpenBMI-MI, HyBRA obtains a comparable BAcc to CSBrain (81.17% versus 81.04%) and improves AUROC and AUPRC by 1.84 and 5.23 percentage points (Page 5).
The performance gain on MentalArithmetic raises the next-best BAcc from 69.64% to 76.63% while using one quarter of the parameters of LaBraM (Page 5).

Ablation and Granularity Analysis

Removing coordinate encoding lowers BAcc by 1.67 percentage points on TUAB and 5.85 points on TUEV (Page 6).
Replacing hyperbolic matching with parameter-matched Euclidean or cosine variants lowers mean BAcc on both datasets (Page 3).
Removing channel-evidence injection reduces BAcc by 1.01 and 6.72 points, supporting the complementary evidence pathway (Page 7).
The complete Region-Time Transformer performs best, showing that temporal modeling within each region and interaction across regions provide complementary information (Page 7).

Conclusion

HyBRA achieves the highest mean BAcc across six downstream datasets by combining a montage-specific region prior with feature-dependent hyperbolic distance, while channelevidence injection preserves information from the original channel tokens (Page 2). The Euclidean and cosine variants on TUAB and TUEV further support the contribution of hyperbolic channel-to-region matching under the evaluated settings (Page 2). Future work will examine adaptive region construction and broader pretraining across centers and electrode configurations (Page 7).

--- Page 3 ---

Coordinate-Aware Patch Encoding Given a batch of EEG segments X ∈ R B×C×T, each channel is divided into P temporal patches and processed by a shared temporal encoder: Hb,c,p = fθ X p b c p, H ∈ R B×C×P ×D, (1) where D is the feature dimension. We encode the threedimensional electrode coordinate qc ∈ R 3 obtained from the standard 10–05 electrode system using ϕ: R 3 → R D and add it to every patch from channel c: He b c p = Hb,c,p + ϕ (qc). (2) The coordinate-aware channel tokens He are passed to the Hyperbolic Region Mixer. More details are in Appendix A.

--- Page 4 ---

Hyperbolic Region Mixer Given coordinate-aware channel tokens He, the Hyperbolic Region Mixer maps montage-dependent observations to R ∈ R B×K×P ×D over a fixed set of K scalp-electrode regions. This shared region axis supports inter-region modeling across montages. Let A ∈ [0, 1]C×K denote the montage-specific channelto-region prior. For a nonempty region k, Ck = C c Ac,k > 0 contains its candidate channels. We map the channel tokens and learnable region prototypes to the Poincaré ball: zb c p = expκ 0 Whheb,c,p + bh, ρk = expκ 0 (µk), db c k p = d κ B (zb c p, ρk). (3)

--- Page 5 ---

Coordinate-Aware Patch Encoding Hyperbolic Region Mixer Region-Time Transformer Multi-scale CNN k=3 k=7 k=15 k=31 x y Coordinate Embedding Temporal Attention Region Attention FFN 4-Layer Channel-to-region prior Scalp-Electrode Regions Beta Activity Delta Activity Sleep Spindle

--- Page 6 ---

Module Setup TUAB TUEV Coordinate-Aware Patch Encoding w/o coordinate encoding 0.8057 ± 0.0051 Hyperbolic Region Mixer Euclidean distance-based matching 0.8134 ± 0.0078 Cosine similarity-based matching 0.8164 ± 0.0214 w/o brain-region tokenization 0.7980 ± 0.027 Hyperbolic Region Mixer w/o channel evidence injection 1 (Page 3)

--- Page 7 ---

Region-Time Transformer The Hyperbolic Region Mixer outputs region-time tokens R ∈ R B×K×P ×D. The Region-Time Transformer operates on this grid. Temporal attention first updates the P patches within each region. Region attention then exchanges information among the K regions at each patch. Learnable region and temporal embeddings are added before the first block: Z0 b k p = Rb k p + e r k + e t p, (9) where e r k, e t p ∈ R D. In the l-th block, temporal attention is followed by region attention.

Improvements for AI systems

To improve transferability across heterogeneous datasets, implement HyBRA to map heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic distances, ensuring that a representation tied to the source channels may entangle task-relevant neural patterns with dataset-specific spatial regularities.
To enhance performance on challenging datasets like OpenBMI-MI, utilize the complementary channel-evidence pathway which draws on pre-aggregation channel features and inject[s] weighted channel-level evidence into each region token, thus addressing the limitation that many-to-one aggregation can attenuate channel-specific evidence.
To achieve superior feature discrimination during aggregation, employ the mechanism where matching weights are calculated as: alpha b,c,k,p = softmax c∈Ck lc,k − lambda db,c,k,p, allowing the system to adapt relative weights based on feature-dependent hyperbolic distance rather than uniform pooling.
To refine spatial representation for BCI tasks like motor imagery and mental arithmetic on TUAB and TUEV, integrate brain-region tokenization using an 11-region partition which the results show yields the highest balanced accuracy, as evidenced by the finding that the 11-region partition further increases BAcc to 81.17 ± 4.36%.
To improve interpretability of model predictions, utilize Region Attribution and Perturbation analysis, which shows that model predictions are more sensitive to highly attributed region tokens, allowing researchers to identify which specific scalp regions contribute most significantly to a prediction by observing the effect of replacing them with fold-specific training-set means.