Mastering ML practices and operations
With the continuous development of large scale software systems, different forms of requirements must be related, in a consistent way, to other activities and artifacts of the development and operational life cycle (dev+ops). We study in this project the improvement of practices in software and Machine Learning (ML) engineering. We propose sound techniques to choose the best ML workflows when contradictions appear (in time series), and to organize these workflows in portofolios. We also study the elicitation and automatic production of requirements justification in software-based experiments and devops contexts. Finally we also investigate new approach for MLOps, i.e., new development principles to change the way AI models are created, deployed, and maintained on the long run.
Ongoing and recent projects:
- Devising new CI/CD principles and techniques to stabilize and improve AI models and especially ML models.
- Reusing experiments and capitalizing knowledge in jupyter notebooks
- Mastering anomalies in machine learning workflow (anomaly detection in time series): what are the requirements that drive the choice of the best workflow? How to learn from apparent contradictions?
- Machine learning workflows portfolio: problem identification through software product lines techniques and meta-learning approaches: http://rockflows.i3s.unice.fr/
- Justification Factory: from justification requirements elicitation to their continuous production
People
Mireille Blay-Fornarino
Professor, (Université Côte d'Azur)
Philippe Collet
Professor, (Université Côte d'Azur), group leader
Yassine Elamraoui
MSc 2019
Yann Brault
MSc 2023
Fredéric Precioso
Professor, (Université Côte d'Azur), MAASAI INRIA-I3S join team