Research map / Architectures and efficiency / Neural architecture search
DC-NAS: Divide-and-Conquer Neural Architecture Search for Multi-Modal Classification
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 DC-NAS, an evolutionary neural architecture search for multi-modal classification that splits the population into sub-populations trained on k-fold stratified subsets (plus one on full data) and exchanges knowledge via two knowledge bases, cutting search time and parameters.
Problem
Neural architecture search for multi-modal classification is very time consuming because evolutionary methods must train and evaluate many full multi-modal networks.
Approach
Divide-and-conquer evolution: most sub-populations evolve on small stratified-fold subsets of the training data while one evolves on the full data, with crossover-based knowledge exchange through two knowledge bases to offset the performance loss from partial-data training; individuals are tree-encoded fusions of unimodal features using five fusion operators.
Key results
- DC-NAS reaches 90.85% cross-subject accuracy on NTU RGB-D (video + pose), vs 90.48% for BM-NAS.
- DC-NAS reaches 95.22% accuracy on EgoGesture (RGB + depth), vs 94.96% for BM-NAS, with 0.19M parameters vs 0.61M.
- On MM-IMDB, DC-NAS reaches 63.70 weighted F1 vs 62.92 for BM-NAS; search time on NTU is 13.63 GPU hours vs 53.68 for BM-NAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DC-NAS instead of BM-NAS | multi-modal classification · MM-IMDB | weighted F1 (%): 62.92±0.03 -> 63.70±0.11 | Helped |
| Use DC-NAS instead of BM-NAS | multi-modal action recognition · NTU RGB-D | accuracy (%): 90.48±0.24 -> 90.85±0.05 | Helped |
| Use DC-NAS instead of BM-NAS | multi-modal gesture recognition · EgoGesture | accuracy (%): 94.96±0.07 -> 95.22±0.05 | Helped |
| Use DC-NAS instead of BM-NAS | architecture search · NTU RGB-D | search cost (GPU hours): 53.68 -> 13.63 | Helped |
| Use DC-NAS instead of BM-NAS | architecture search · EgoGesture | search cost (GPU hours): 20.67 -> 4.57 | Helped |
| Use DC-NAS instead of BM-NAS | fusion model · NTU RGB-D | model parameters: 0.98M -> 0.26M | Helped |
| Train all individuals on the entire training dataset instead of dividing the population and data (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85±0.05 -> 90.86±0.03 | No clear effect |
| Remove knowledge transfer (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85±0.05 -> 90.52±0.06 | Hurt |
| Use random feature selection instead of searched feature selection (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85±0.05 -> 88.81±0.11 | Hurt |
| Use late fusion instead of searched feature selection (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85±0.05 -> 89.47±0.07 | Hurt |
| Use addition as the fusion operator instead of the searched fusion strategy (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85 -> 89.54 | Hurt |
| Use multiplication as the fusion operator instead of the searched fusion strategy (ablation) | DC-NAS · NTU RGB-D | accuracy (%): 90.85 -> 88.71 | Hurt |
Surprising findings
- The full-data-only variant (DC-NAS1) is not significantly better than DC-NAS, while removing knowledge transfer (DC-NAS2) lowers accuracy to 90.52%, though the time saving of the partial-data variant is smaller than expected (11.10 vs 13.63 hours for DC-NAS2 vs DC-NAS).
Limitations
- Gains over BM-NAS are small (e.g., 0.37 points on NTU) and rely on repeated runs with modest significance.
- Knowledge exchange and data-split strategies are simple and acknowledged as needing deeper study.
- Evaluated only on three tasks with two modalities each, with backbones fixed to those used by BM-NAS.
Experimental setup
Datasets: MM-IMDB NTU RGB-D EgoGesture
Benchmarks: MM-IMDB NTU RGB-D EgoGesture
Metrics: weighted F1 accuracy parameter count search time (GPU hours) paired t-test p-value
Baselines: MFAS BM-NAS MMTM CentralNet GMU Two-stream MTUT 3D-CDC-NAS
Builds on: MFAS BM-NAS MMTM island-based evolutionary algorithms
Closest papers by meaning
- MFAS: Multimodal Fusion Architecture Search (CVPR 2019 · similarity 0.86)
- BM-NAS: Bilevel Multimodal Neural Architecture Search (AAAI 2022 · similarity 0.85)
- Runtime Analysis of Evolutionary NAS for Multiclass Classification (ICML 2025 · similarity 0.85)
- Neural Fine-Tuning Search for Few-Shot Learning (ICLR 2024 · similarity 0.84)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.83)
- MOTE-NAS: Multi-Objective Training-based Estimate for Efficient Neural Architecture Search (NeurIPS 2024 · 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.