Research map / Label-efficient and robust learning / Recognition and long-tail
DARNet: Dual Attention Refinement Network with Spatiotemporal Construction for Auditory Attention Detection
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 DARNet, a compact EEG auditory attention detection network combining CSP-based spatiotemporal convolutions, two stacked attention-refinement blocks, and multi-level feature fusion, reporting SOTA accuracy with ~0.08M parameters on KUL, DTU and MM-AAD.
Problem
Auditory attention detection from EEG struggles to use spatial distribution information and to capture long-range temporal dependencies, and existing models are large and need long decision windows.
Approach
Combine a temporal and a full-channel spatial convolution on CSP-filtered EEG with two stacked multi-head self-attention plus convolution-max-pool refinement layers, fusing shallow and deep attention features for classification.
Key results
- DARNet reaches 96.2% on KUL audio-only at 1-second window and 87.8% on DTU at 1-second window, versus DBPNet's 95.0% and 83.9%.
- At 0.1-second window, DARNet achieves 79.5% on DTU (DBPNet 75.1%) and 91.6% on KUL (DBPNet 87.1%).
- DARNet has 0.08M trainable parameters versus 0.91M for DBPNet, a reduction of about 91%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatial feature extraction (ablation) | DARNet · DTU audio-only, 0.1-second | accuracy (%): 71.5 \(\pm\) 5.79 -> 79.5 \(\pm\) 5.84 | Helped |
| Add temporal feature extraction (ablation) | DARNet · DTU audio-only, 0.1-second | accuracy (%): 70.9 \(\pm\) 5.62 -> 79.5 \(\pm\) 5.84 | Helped |
| Add feature fusion module (ablation) | DARNet · DTU audio-only, 0.1-second | accuracy (%): 77.5 \(\pm\) 7.07 -> 79.5 \(\pm\) 5.84 | Helped |
| Use two attention refinement layers instead of one (ablation) | DARNet · DTU audio-only, 1-second | accuracy (%): 86.3 \(\pm\) 5.83 -> 87.8 \(\pm\) 6.02 | Helped |
| Use two attention refinement layers instead of one (ablation) | DARNet · KUL audio-only, 1-second | accuracy (%): 95.5 \(\pm\) 3.28 -> 96.2 \(\pm\) 3.04 | Helped |
| Fit CSP after dataset splitting to avoid leakage | DARNet · KUL audio-only, 0.1-second | accuracy (%): 91.6 \(\pm\) 4.83 -> 89.2 \(\pm\) 5.50 | Hurt |
| Increase decision window from 0.1-second to 1-second | DARNet · KUL audio-only | accuracy (%): 91.6 -> 96.2 | Helped |
| Use audio-visual rather than audio-only input | DARNet · MM-AAD, 0.1-second | accuracy (%): 94.9 -> 95.8 | Helped |
Surprising findings
- After correcting a data leakage issue (CSP fit before splitting), DARNet's accuracy drops notably, e.g. DTU 1-second from 87.8% to 80.1%, and falls below corrected DBPNet on DTU and MM-AAD in some settings.
Limitations
- Evaluated only in subject-dependent setting; cross-subject generalization not tested.
- Baseline results for several methods are copied from prior papers rather than re-run.
- No statistical significance tests or compute resource details reported.
Experimental setup
Datasets: KUL DTU MM-AAD
Benchmarks: KUL DTU MM-AAD
Metrics: classification accuracy standard deviation parameter count
Baselines: SSF-CNN MBSSFCC BSAnet DenseNet-3D DBPNet EEG-Graph Net
Builds on: DBPNet common spatial patterns Informer-style self-attention distilling
Closest papers by meaning
- S$^2$M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection (NeurIPS 2025 · similarity 0.85)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.83)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.83)
- Neural decoding from stereotactic EEG: accounting for electrode variability across subjects (NeurIPS 2024 · similarity 0.81)
- LAtte: Hyperbolic Lorentz Attention for Joint-Subject EEG Classification (NeurIPS 2026 · similarity 0.81)
- M3Net: Efficient Time-Frequency Integration Network with Mirror Attention for Audio Classification on Edge (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.