For the past five years, Michelin has been working on incorporating more machine learning into its processes for tire design and testing. Léo Dreyfus-Schmidt, VP of Research at Dataiku and head of the AI Lab, sat down with François Deheeger, Senior Fellow AI and Data Science at Michelin to talk details.READ ON THE BLOG
Dataiku AI Lab
The gap between machine learning research and how today's organizations are actually leveraging machine learning can be wide. Enter: the Dataiku AI Lab. Our team of experts builds bridges between machine learning research and practical business applications to accelerate our customers' journey to Everyday AI.
What Is the Dataiku AI Lab?
The Dataiku AI Lab is a team of researchers that develops robust and generic state-of-the-art machine learning methods to enable trustworthy, real-life applications. The team has a broad range of interests, from active learning to uncertainty estimation, data shifts to casual ML, and more. Learn about a few of the team’s current projects below.
Current Research: Addressing Dataset Shift
Machine learning model lifecycles are complex and ever evolving. When data changes or drifts, whether naturally or adversarially, it can impact a model’s performance. We need to not only detect data drift but also estimate model performance drop. Check out Dataiku AI Lab’s latest papers on this topic:
- Ensembling Shift Detectors: an Extensive Empirical Evaluation – ECML PKDD 2021
- Performance Prediction Under Dataset Shift – ICPR 2022
Current Research: Exploring Active Learning
When unlabeled data is abundant but labeling resources are limited, active learning leverages knowledge from the model to select samples that maximize the model performance. Dataiku AI Lab has published the following papers on active learning:
- Rebuilding trust in active learning with actionable metrics – ICDM 2020
- Sample Noise Impact on Active Learning – IAL workshop at ECML-PKDD 2021
- Cardinal, a metric-based Active learning framework – SIMPAC journal 2022
- OpenAL: Evaluation and Interpretation of Active Learning Strategies – NeurIPS
- Human in the Loop Workshop 2022.
We also open source our Python package, Cardinal.