HEEGNet: Hyperbolic Embeddings for EEG
Shanglin Li, Shiwen Chu, Okan Ko\c{c}, Yi Ding, Qibin Zhao, Motoaki Kawanabe, Ziheng Chen
TL;DR
This paper introduces HEEGNet, a hybrid hyperbolic neural network that leverages hyperbolic embeddings to better capture hierarchical structures in EEG data, improving generalization across domains in brain-computer interfaces.
Contribution
The study demonstrates EEG data's hyperbolicity and proposes HEEGNet, a novel hybrid hyperbolic network with a domain adaptation strategy for improved EEG decoding.
Findings
Hyperbolic embeddings enhance EEG data representation.
HEEGNet achieves state-of-the-art results on multiple EEG datasets.
The approach improves generalization across different subjects and tasks.
Abstract
Electroencephalography (EEG)-based brain-computer interfaces facilitate direct communication with a computer, enabling promising applications in human-computer interactions. However, their utility is currently limited because EEG decoding often suffers from poor generalization due to distribution shifts across domains (e.g., subjects). Learning robust representations that capture underlying task-relevant information would mitigate these shifts and improve generalization. One promising approach is to exploit the underlying hierarchical structure in EEG, as recent studies suggest that hierarchical cognitive processes, such as visual processing, can be encoded in EEG. While many decoding methods still rely on Euclidean embeddings, recent work has begun exploring hyperbolic geometry for EEG. Hyperbolic spaces, regarded as the continuous analogue of tree structures, provide a natural…
Peer Reviews
Decision·ICLR 2026 Poster
1. The paper tackles an important issue—cross-domain generalization in EEG signals—where conventional Euclidean-based networks often fail to model hierarchical dependencies or inter-domain variations effectively. 2. The hybrid use of Euclidean and hyperbolic embeddings provides an interesting perspective, allowing the model to balance local feature encoding and hierarchical structural representation. This design is conceptually meaningful for EEG data, which may contain both flat temporal dynami
1. The manuscript claims that EEG, video, and language modalities can all be represented in hyperbolic space, but does not provide sufficient theoretical or empirical justification for why EEG data, in particular, exhibits hierarchical or tree-like properties that make hyperbolic geometry appropriate. 2. The rationale for combining Euclidean and hyperbolic representations is underdeveloped. It remains unclear what complementary features each space captures and how their joint use specifically be
The two-step alignment strategy, DSMDBN, operates in hyperbolic space: it first aligns source and target domain distributions by matching their first and second moments using a hyperbolic BatchNorm module, and then further aligns the moment-normalized features to a standard hyperbolic Gaussian via the Horospherical Sliced-Wasserstein (HHSW) loss. This approach is well-motivated and offers a moderately novel contribution to the domain adaptation literature.
1. Hyperbolicity evidence may reflect model bias: The observed hyperbolicity is derived from learned embeddings rather than the intrinsic geometry of the EEG signals. Therfore, the results may reflect model-induced structure rather than inherent data-level hyperbolic properties. 2. Numerical stabilization not addressed: The paper does not discuss numerical stabilization techniques essential for reliable hyperbolic training, such as gradient clipping, norm normalization, or overflow prevention
1. The research direction of using hyperbolic embeddings for EEG-based BCIs/recognitions is emerging and promising. 2. The proposed method is well explained. 3. The performance improvements compared with EEGNet are significant.
1. The novelty is not well clarified; and the idea of exploring hyperbolic space /embeddings for EEG-based recognition is not new. The authors didn't clearly explain the difference. 2. My main concern is about the performance comparison part. -- The proposed HEEGNet was not compared with other hyperbolic embeddings based methods, e.g., Jing Chang 2025 in the ref list and related refs/citations of that paper. -- For different tasks/datasets, it doesn't seem that the proposed methods were compa
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEEG and Brain-Computer Interfaces · Emotion and Mood Recognition · Functional Brain Connectivity Studies