Driverless AI: Your Expert System for AI
- Observability
- H2O Driverless AI
Overview
H2O.ai CEO Sri Ambati and CTO Arno Candel introduce a major Driverless AI release focused on GPU-accelerated speed, Kaggle Grandmaster-built recipes, and a new machine learning interpretability (MLI) suite. The session covers moving gradient boosting and clustering algorithms onto GPUs for large speedups, then a live demo of MLI's surrogate models, reason codes, and automatic English explanations for a credit-default model. The team also walks through Docker-based installation and takes audience questions on clustering, licensing, and Kaggle usage.
Key takeaways
- How H2O.ai moved gradient boosting, GLM, and k-means clustering onto GPUs, with speedups shown live (up to 40x on a GLM example)
- Why Kaggle Grandmaster techniques like target encoding are built into Driverless AI as automated recipes, and the leakage risk if they're done incorrectly
- How the machine learning interpretability (MLI) suite explains a live credit-default model using a surrogate decision tree, reason codes for individual predictions, and automatically generated English explanations
- How the K-LIME and partial dependence views let you compare a single prediction against the global population to see which features drove it
- How to install Driverless AI as an Nvidia Docker image, including the setup steps and run command
- Answers to audience questions on multi-GPU limits, whether the tool works on Kaggle, and which parts of the stack stay open source (Datatable, H2O GPU Edition)
Start owning your AI