Reproduce “Evolutionary Neural Architecture Search for 3D Point Cloud Analysis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yisheng Yang, Guodong Du, Chean Khim Toa, Ho-Kin Tang, Sim Kuan Goh
SHSADE-PIDS, an evolutionary Neural Architecture Search framework, efficiently designs deep neural networks for 3D point cloud analysis. This framework discovers compact architectures that achieve up to 64.51% mIoU on SemanticKITTI and 93.4% accuracy on ModelNet40, while reducing parameters by up to 34 times and MACs by up to 14 times compared to prior methods.
Key insights
- Adaptive evolutionary algorithms, like the proposed SHSADE, can effectively bridge continuous optimization with discrete neural architecture search, leading to improved convergence and search efficiency.
- Jointly optimizing point interactions and architectural dimensions within an evolutionary NAS framework can discover highly compact yet accurate models for challenging 3D point cloud tasks.
- The use of a fast predictive model for evaluating candidate architectures is crucial for making the evolutionary search process computationally feasible for complex 3D architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.