Microsoft Project Petridish: Efficient Forward Neural Architecture Search Cuts Training Costs

It's a research project introducing a new neural architecture search (NAS) method based on an efficient forward approach, evaluating candidates while training a single backbone model, reducing costs.

While ENAS relies on weight sharing, Petridish outperforms it by achieving 20% higher efficiency in reducing training costs, with similar or better accuracy.

Yes, the code is open-source, and technical teams in the region can experiment with it to accelerate custom model development at lower costs, especially in sectors like banking and healthcare.