Panel: OpML
summary
Operationalizing ML faces subtle failures and distribution shifts that are hard to detect — a schema drift takes three weeks to diagnose in one recommendation pipeline, a 3% geographic data drop wrecks a model for a language segment. Practical fixes include programmable golden set checks, layered anomaly detection, and investing in model quality monitoring. Building a single common ML platform across an organization (an 18‑month migration at one firm) speeds rollout and bias fixes, but lifecycle to GA often spans ~2 years and governance and infra outages remain primary risks.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 1:30Panel Introductions
- 3:40Key differences between OpML and other platforms
- 9:32What observability looks like
- 15:28Life cycle management
- 15:44Life cycle of OpML
- 17:08Organizational structure
- 18:15Immature industry
- 20:53Transfer learning
- 22:03Change one thing at a time
- 24:05Sharing experiences
- 24:31Distributed systems failures
- 26:49Machine learning teams
- 27:38Applications of ML
- 27:53Use Cases
- 29:34Where do people put the ML
- 32:00Governance
- 33:52Regulations
- 34:41Fraud
- 37:10Magic
- 37:52Platforms
- 39:39Azure Machine Learning
- 41:01AWS vs GCP
- 42:28ML platforms
- 45:03Outro
related talks
42mDeploying ML models and all the things that go wrong
30mChallenges of machine learning: The new ML lifecycle
19mHow ML Breaks: A Decade of Outages for One Large ML Pipeline
32mTeam Teardown: Evolving an ML Platform
13mMLOps: From AI Research to Production-Ready Systems
32mMaking Models Work: How an optimized ML lifecycle can drive better ROI
41mML plumbing
34mScaling ML in the Traditional Enterprise