Reproduce “Manas: Mining Software Repositories to Assist AutoML” 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
Giang Nguyen, Md Johir Islam, Rangeet Pan, Hridesh Rajan
Manas introduces a method that enhances Neural Architecture Search (NAS) by mining CNN architectures and optimizers from open-source repositories. This approach generates simpler, faster-training models with lower error rates on most evaluated Kaggle datasets, achieving an average of 75.7% less depth and 93.0% fewer parameters for image classification compared to Auto-Keras.
Key insights
- Integrating knowledge from human-developed models in software repositories can significantly enhance NAS efficiency by providing more suitable starting points and informed architectural modifications.
- Specific architectural transformations (e.g., Batch Normalization, Global Average Pooling, Dropout) and the use of developer-selected optimizers improve model convergence and reduce error rates during NAS.
- Efficient model matching through clustering and filtering can drastically reduce the number of candidate initial architectures, allowing for more effective allocation of computational resources during the search.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.