Reproduce “Generative flow induced neural architecture search: Towards discovering optimal architecture in wavelet neural operator” 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
Hartej Soin, Tapas Tripura, Souvik Chakraborty
Generative Flow Induced Neural Architecture Search (GFN-NAS) was developed at the Indian Institute of Technology Delhi to automate the discovery of optimal Wavelet Neural Operator (WNO) architectures for solving partial differential equations. This approach consistently enhanced WNO prediction accuracy across various fluid dynamics problems by identifying layer-specific wavelet bases and activation functions.
Key insights
- Optimal neural operator architectures for scientific computing problems often consist of heterogeneous, layer-specific combinations of components (e.g., different wavelet bases and activation functions per block).
- GFlowNets provide an efficient framework for NAS by learning to sample diverse, high-reward architectures proportionally, circumventing the extensive episodic training typically required by reinforcement learning-based NAS.
- Evaluating partially trained neural operators is sufficient to provide effective reward signals for guiding the architecture search process, significantly improving computational efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.