Research map / Label-efficient and robust learning / Recognition and long-tail

LAtte: Hyperbolic Lorentz Attention for Joint-Subject EEG Classification

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LAtte, a fully hyperbolic (Lorentz) InceptionTime-based attention model with subject-specific boost-based LoRA adapters for joint-subject EEG classification, evaluated in subject-specific, subject-conditional and LOSO settings on five EEG datasets.

Problem

EEG classification suffers from low signal-to-noise ratio and high inter-subject variability; existing methods mostly train subject-specific models that do not exploit shared structure or generalize to unseen subjects.

Approach

Combines a Lorentz (hyperboloid) InceptionTime encoder with a dual-branch baseline-deviation design (average-pool baseline minus max-pool task branch), Lorentz attention, and subject-specific Lorentz boost LoRA adapters plus a frozen random-projection prototype decoder; a subject-ID masking scheme (slot 0) enables LOSO generalization.

Key results

• Subject-specific with fine-tuning: LAtte+FT reaches 79.09% on BCIC-IV 2a vs 74.71% for MAtt, and 73.47% on MAMEM II vs 68.10% for HyperMAtt (reported relative improvements of 5.86% and 8.30% in the table).
• Subject-conditional LAtte achieves 78.85% on BCIC-IV 2a and 83.8 AUC on BCI Challenge, with reported relative gains of 28.46% and 10.07% over joint baselines.
• LOSO on BCIC-IV 2a reaches 65.32% vs 63.00% for TCFormer (3.68% relative improvement); on MAMEM II LOSO LAtte achieves 57.86% vs 50.31% for ATCNet.

Surprising findings

• Foundation model CBraMod underperforms small-scale EEG baselines on these datasets (e.g., 45.22% on BCIC-IV 2a SS), and TCFormer nearly fails on several MAMEM II subjects under LOSO.
• Raw BCI Challenge and BCIC-IV 2a signals were less hyperbolic than shuffled baselines, becoming hyperbolic only after the encoder input block.

Limitations

• Generalization across datasets is not tested and requires architectural changes for varied input lengths.
• Hyperbolic operations lack optimized CUDA implementations, making them slower per operation.
• Gains on the large HGD dataset are within margin of error relative to TCFormer.

Experimental setup

Datasets: BCIC-IV 2a BCIC-IV 2b MAMEM II BCI Challenge High Gamma Dataset

Benchmarks: BCIC-IV 2a BCIC-IV 2b MAMEM II BCI Challenge High Gamma Dataset

Metrics: accuracy AUC

Baselines: ShallowConvNet EEGNet SCCNet EEG-TCNet TCNet-Fusion FBCNet MBEEGSE InceptionTime MAtt CBraMod HyperMAtt EEGFormer

Builds on: MAtt HyperMAtt InceptionTime TCFormer LoRA Lorentz prototype classifier

Closest papers by meaning

• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.86)
• SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG (ICLR 2025 · similarity 0.86)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.85)
• MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding (NeurIPS 2026 · similarity 0.84)
• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.84)
• REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects (NeurIPS 2025 · similarity 0.84)