Continuous ML Applications in Production
summary
PipelineAI shows a production runtime that continuously trains, optimizes, validates and explains models using Continuous VOTE techniques, feeding online metrics into mlflow and supporting Kafka streams, Docker on Kubernetes and TensorFlow/Keras/MXNet/XGBoost runtimes. Fewer than 5% of offline models reach production and model variants take about one to three months to test, so the system runs shadow mode comparisons, online hyperparameter tuning, ensemble routing, cost-per-prediction estimates and cross-cloud traffic shifting to pick the fastest or cheapest variant.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:23AI runtime concept explained
- 4:16Batch vs continuous ML
- 8:12Problem with model handoff
- 10:23Modern frameworks and hardware
- 14:58Demo setup and features
- 17:53Auto-scaling and routing
- 20:59Q&A on online training
related talks
59mContinuous Intelligence: Keeping your AI Application in Production
01h 05mMachine Learning Models and Productionization
13mMLOps: From AI Research to Production-Ready Systems
47mMachine Learning in Production: An Intro to MLOps.
42mDeep learning beyond the learning
17mContinuous deployment for ML: The new software development lifecycle
32mAutomating Production Production Level Machine Learning on AWS
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD