Analyze a Completed Hydrogen Torch Experiment
- Predictive AI
- H2O Hydrogen Torch
This lesson walks through reviewing a completed experiment in H2O Hydrogen Torch, starting with locating the final validation metric once an experiment's status changes to finished. It covers using the Validation prediction insights tab to examine the best and worst validation samples, revealing that dark or unclear images produced the highest loss values. It then compares two use cases for the same coin counting model, showing a validation MAE too high for an ATM's exact counting needs but acceptable for a rapid approximate sum. It closes by tying model acceptability back to the accuracy requirements of the intended use case.
Key takeaways
- How to find an experiment's final validation metric once its status changes to finished in Hydrogen Torch
- What the best and worst validation samples reveal, in this case that dark images caused the highest loss
- Why the same validation MAE can be unacceptable for exact ATM coin counting but fine for a rough estimate
- How reviewing worst performing samples can point to a data fix, like replacing unclear images
- The acceptability of a model depends on matching its error margin to the specific use case requirements
Start owning your AI