Research map / Architectures and efficiency / Neural architecture search
BM-NAS: Bilevel Multimodal Neural Architecture Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Search both feature selection and fusion structure with bilevel BM-NAS | Multimodal fusion models · MM-IMDB | Weighted F1 (%): 62.50 -> 62.92 $\pm$ 0.03 | Helped |
| Search both feature selection and fusion structure with bilevel BM-NAS | Multimodal fusion models · NTU RGB-D | Accuracy (%): 89.50 $\pm$ 0.60 -> 90.48 $\pm$ 0.24 | Helped |
| Search both feature selection and fusion structure with bilevel BM-NAS | Multimodal fusion models · EgoGesture | Accuracy (%): 94.38 -> 94.96 $\pm$ 0.07 | Helped |
| Use BM-NAS instead of MFAS for architecture search | Multimodal fusion search · MM-IMDB | Search cost (GPU·hours): 9.24 -> 0.89 | Helped |
| Use BM-NAS instead of MFAS for architecture search | Multimodal fusion search · NTU RGB-D | Search cost (GPU·hours): 603.64 -> 38.6 | Helped |
| Replace random feature and connection selection with searched feature selection (ablation) | BM-NAS on NTU RGB-D · NTU RGB-D | Accuracy (%): 86.35 $\pm$ 0.68 -> 90.48 $\pm$ 0.24 | Helped |
| Replace late fusion with searched feature selection (ablation) | BM-NAS on NTU RGB-D · NTU RGB-D | Accuracy (%): 89.49 $\pm$ 0.15 -> 90.48 $\pm$ 0.24 | Helped |
| Replace MFAS feature selection with BM-NAS-searched feature selection (ablation) | BM-NAS on NTU RGB-D · NTU RGB-D | Accuracy (%): 89.50 $\pm$ 0.60 -> 90.48 $\pm$ 0.24 | Helped |
| Replace summation with BM-NAS-searched fusion (ablation) | BM-NAS with fixed feature selection · NTU RGB-D | Accuracy (%): 87.64 -> 90.48 | Helped |
| Replace ConcatFC with BM-NAS-searched fusion (ablation) | BM-NAS with fixed feature selection · NTU RGB-D | Accuracy (%): 89.20 -> 90.48 | Helped |
| Replace multi-head attention with BM-NAS-searched fusion (ablation) | BM-NAS with fixed feature selection · NTU RGB-D | Accuracy (%): 88.29 -> 90.48 | Helped |
| Replace Attention on Attention with BM-NAS-searched fusion (ablation) | BM-NAS with fixed feature selection · NTU RGB-D | Accuracy (%): 89.11 -> 90.48 | Helped |
Closest papers by meaning
- DC-NAS: Divide-and-Conquer Neural Architecture Search for Multi-Modal Classification (AAAI 2024 · similarity 0.85)
- Multimodal Graph Neural Architecture Search under Distribution Shifts (AAAI 2024 · similarity 0.84)
- L-SWAG: Layer-Sample Wise Activation with Gradients Information for Zero-Shot NAS on Vision Transformers (CVPR 2025 · similarity 0.84)
- RZ-NAS: Enhancing LLM-guided Neural Architecture Search via Reflective Zero-Cost Strategy (ICML 2025 · similarity 0.83)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.83)
- Fusion or Confusion? Multimodal Complexity Is Not All You Need (NeurIPS 2026 · similarity 0.83)
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.