Summary
In this chapter, we have implemented NAS, a framework where a reinforcement learning agent (the Controller) generates child neural networks to complete a certain task. We studied the theory behind how the Controller learns to generate better child network architectures via policy gradient methods. We then implemented a simplified version of NAS that generates child networks that learn to classify CIFAR-10 images.
For more information on related topics, refer to the following list of links:
- NAS with reinforcement learning:Â https://arxiv.org/abs/1611.01578
- Efficient NAS via parameter sharing:Â https://arxiv.org/pdf/1802.03268
- Google Cloud AutoML:Â https://cloud.google.com/automl/
- Awesome Architecture Search—a curated list of papers related to generating neural networks: https://github.com/markdtw/awesome-architecture-search
The NAS framework marks an exciting development in the deep learning field, for we have figured out how to automatically design neural network architectures, a decision...
