Neural Architecture Search (NAS) automates the process of architecture design of neural networks. NAS approaches optimize the topology of the networks, incl. how to connect nodes and which operators to choose. User-defined optimization metrics can thereby include accuracy, model size or inference time to arrive at an optimal architecture for specific applications. Due to the extremely large search space, traditional evolution or reinforcement learning-based AutoML algorithms tend to be computationally expensive. Hence recent research on the topic has focused on exploring more efficient ways for NAS. In particular, recently developed gradient-based and multi-fidelity methods have provided a promising path and boosted research in these directions. Our group has been very active in developing state of the art NAS methods and has been at the forefront of driving NAS research forward. We give a summary of a few recent important work released from our group –
NASLib is a modular and flexible Neural Architecture Search (NAS) library. Its purpose is to facilitate NAS research in the community and allow for fair comparisons of diverse recent NAS methods by providing a common modular, flexible and extensible codebase. The library also provides an easy-to-use interface with the popular NAS benchmarks (e.g. NASBench 101, 201, 301), with various discrete (e.g. Random Search, Regularized Evolution) and one-shot (DARTS, GDAS, DrNAS, etc.) optimizers already integrated in.
The library currently has an active, strong, developer community who are maintaining it along with constantly adding new benchmarks, optimizers and search spaces into the code base.
Based on the well-known DL framework PyTorch, Auto-PyTorch automatically optimizes both the neural architecture and the hyperparameter configuration. To this end, Auto-PyTorch combines ideas from efficient multi-fidelity optimization, meta-learning and ensembling.
Research on NAS is often very expensive because training and evaluating a single deep neural network might require between minutes or even days. Therefore, we provide several benchmark packages for NAS that either provide tabular or surrogate benchmarks, allowing efficient research on NAS.
Best practices for NAS Research
The rapid development of new NAS approaches makes it hard to compare these against each other. To ensure reliable and reproducible results, we also provide best practices for scientific research on NAS and our checklist for new NAS papers.
Selected NAS Papers
NAS is one of the booming subfields of AutoML and the number of papers is quickly increasing. To provide a comprehensive overview of the recent trends, we provide the following sources:
- NAS survey paper [JMLR 2020]
- A book chapter on NAS from our open-access book, “AutoML: Methods, System, Challengers”
- A continuously updated page with a comprehensive NAS literature overview
One-Shot NAS Methods
- Understanding and Robustifying Differentiable Architecture Search [ICLR 2020, Oral]
- Meta-Learning of Neural Architectures for Few-Shot Learning [CVPR 2020]
Neural Ensemble Search
- Neural Ensemble Search for Performant and Calibrated Predictions [ICML 2020, UDL Workshop Oral]
Joint NAS and Hyperparameter Optimization
- Towards Automated Deep Learning: Efficient Joint Neural Architecture and Hyperparameter Search [ICML 2018, AutoML Workshop]