Using rewrites we can produce many versions of the same model very easily in Aesara. For instance we could automatically look for the kernel of a Gaussian process, change the structure of neural networks, etc. In deep learning this called Neural Architecture Search. Rewrites implement facts.

We need to find a simple example.

  • Vanilla Gaussian Process (GPae)
  • Simples neural network architecture (Kaeras)
  • Neural Architecture Search with Reinforcement Learning (ArxiV)
  • Structure Discovery in Nonparametric Regression Through Compositional Search (ArXiv)