The Structure and Interpretation of ML Metadata
summary
Claims that ML metadata is the development tool that actually makes software 2.0 reproducible and governable, demonstrated by the Aim open source experiment tracker used in a good number of large companies. Shows metadata types (dictionaries and sequences), tracks hundreds of metrics per run and GPU temperature and driver memory, captures lineage from data preprocessing to production, and connects frameworks and resource managers (Kubernetes Ray PyTorch Lightning) so pipelines and model promotion/monitoring (data drift detection) are reproducible and debuggable.
AI-written summary · about
chapters (13)
- 0:00ML metadata introduction
- 1:14Software 1.0 vs 2.0 in ML
- 3:45Defining ML systems
- 9:06Loops in ML architectures
- 10:30What is ML metadata
- 15:06ML metadata types
- 16:52ML metadata applications
- 18:17ML vs 1.0 metadata differences
- 20:28Why ML metadata matters
- 22:11Advanced ML metadata uses
- 25:43ML metadata as infrastructure glue
- 26:26Future of ML metadata stacks
- 27:09Q&A
related talks
29mMetadata Systems for End-to-End Data & Machine Learning
10mML-based Log Analysis for Faster Debugging
33mDiscover How to Create Your Own Metadata-Driven ML Platform from Scratch
32mReproducibility and versioning of ML systems
28mthe semantic layer your AI needs
41mML plumbing
06mML for Systems and Systems for ML
29mReproducibility in ML Systems: A Netflix Original