Humans helping machines help humans run machines
summary
Benchling builds an automated pipeline that produces usable models on very small, custom tabular growth datasets (far below the ~3,000 rows where deep learning helps). It uses Featuretools for deep feature synthesis, an AutoML step that allows injected domain pipeline choices, and SHAP plus permutation importance for explanations. The system applies grouping cross‑validation for batch effects, uses tree models, exposes a UI for config and in beta interactive what‑if simulations, and pilots show it can scale expertise across customers.
AI-written summary · about
chapters (11)
- 0:00Automation and analytics intro
- 1:26Machine learning in lab work
- 4:17Benchling lab data structure
- 6:25Automated modeling pipelines
- 11:20Models as interactive tools
- 14:36Q&A experiments
- 17:49Q&A clinical trials
- 19:53Q&A customer feedback
- 20:38Q&A lab data types
- 22:04Q&A model explanations
- 23:48Q&A ML patterns
related talks
01h 40mBuilding machine learning pipelines that scale
32mTesting Machine Learning Systems: What it is and why you should care
38mBuilding robust machine learning systems
21mBuilding Scalable ML Systems: Strategies and Insights
18mAutomate the boring ML stuff with pipelines
42mDeep learning beyond the learning
42mImproving Machine Learning Development Reliability
27mHow to make ML teams 80% more productive.