TL;DR
LAtte introduces a hyperbolic Lorentz attention framework with structured EEG signal decomposition and low-rank adaptation, significantly enhancing cross-subject EEG classification performance across multiple datasets.
Contribution
The paper proposes LAtte, a novel EEG classification model combining Lorentz attention, hyperbolic encoding, and low-rank adaptation for improved generalization and robustness.
Findings
Consistently outperforms state-of-the-art methods on five EEG datasets.
Improves performance especially on smaller datasets.
Maintains high accuracy in leave-one-subject-out evaluations.
Abstract
Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain-computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The hyperbolic design extending prior work by incorporating hyperbolic operations in both encoder and decoder. It leverages hyperbolic geometry to better model hierarchical structures in EEG data. LAtte's use of LoRA for subject embeddings allows joint training across subjects while maintaining adaptability. The self-supervised pretraining tasks effectively handle noisy EEG data, and the Johnson-Lindenstrauss-inspired projection adds efficient regularization. Ablation studies demonstrate the
The motivation for hyperbolic space is sound, but analysis of why it suits EEG hierarchies is needed, which is currently missing. The paper lacks discussion on training/inference time, parameter counts, or scalability compared to baselines. Hyperbolic operations can be computationally intensive, so benchmarching could be useful. For sections such as the Lorentz operations and attention mechanism, additional intuitive explanations on motivations behind the design and related equations would be
1. **Novel Architectural Contribution:** Both reviews recognize LAtte as the first fully hyperbolic pipeline for EEG classification, representing a meaningful step beyond prior partially hyperbolic approaches. 2. **Cross-Subject Focus Addresses Important Problem:** The cross-subject generalization motivation is genuine and well-articulated. 3. **Comprehensive Ablation Studies:** The ablation study systematically validates component contributions. 4. **Presentation Has Accessibility Strengths:**
1. **Missing Foundation Model Baselines:** Recent foundation-style EEG models (CebraMod, LaBraM) demonstrate strong cross-subject transfer and state-of-the-art performance Without these comparisons, cannot determine if hyperbolic design advantages outweigh data-driven massive pretraining *Impact:* Undermines competitive positioning claims and makes it unclear whether architectural innovation provides value beyond scale 2. **LOSO Validation Missing:** Standard protocol for validating true cross-s
1. The LoRA-based subject adapters offer a lightweight way to encode subject identity and separate subject-specific noise distributions while still training a single joint model. This addresses a frequent bottleneck in real BCI workflows (per-subject retraining). 2. Explicitly moving beyond single-subject (SS) training to focus on the more clinically relevant subject-conditional (SC) setting is a strong point 3. LAtteJoint reports state-of-the-art results in the SC setting across all three datas
1. The discussion of related work reads more like a list of prior studies rather than a coherent synthesis. The relationships among different works are unclear, and the authors fail to explicitly connect the cited literature to the motivation or design choices of their own model. This section requires substantial restructuring to establish a logical narrative. 2. The innovation of this paper is quite limited. Its components (such as the Lorentz attention mechanism and hyperbolic functional conn
Novel use of fully hyperbolic architecture tailored for EEG.
1. Unsubstantiated Hierarchy Claim The paper makes a strong claim that EEG data has a hierarchical structure inherently suited for hyperbolic space. However, no empirical evidence is provided to support this assertion. References to the spatial arrangement of EEG sensors are insufficient, as sensor layout does not imply a latent hierarchy in the underlying data manifold. 2. Missing Comparisons with Recent Foundation Models This work lacks comparison with recent high-impact foundation models s
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
