Learn why continual learning is important, and how to use it in your machine learning models to improve accuracy. You can download the full webinar here: https://info.cnvrg.io/continual-learning-webinar
2. whoami
• Developer/Data scientist => CEO
• cnvrg.io – built by data scientists, for data scientists to help teams:
• manage, build and automate machine learning
• bridge science and engineering
3. agenda
• Introduction to Continual Learning
• Why do we need continual learning?
• An ML Pipeline with Continual Learning
• AutoML
• Deployment
• Monitoring
• Live Demo
4. Continual learning (CL) is the ability of a model to learn continually from a
stream of data, updating the model in production to maintain performance
and relevancy
16. MNIST ML Pipeline with Continual Learning
- OCR application to identify numbers from handwritten text
- User base is growing, and our model should support new users
17. DATA Data validation AutoML & HPO Model Validations Model Deployment
Predictions
MNIST ML Pipeline with Continual Learning
MonitoringCleaning & Labeling
18. MNIST ML Pipeline with Continual Learning
DATA Data validation AutoML & HPO Model Validations Model Deployment
PredictionsMonitoringCleaning & Labeling
19. • Data is changing -> Models are changing
• Select your algorithm space
• For example, for vision: VGG/Inception/ResNet
• For each algorithm, specify its range of hyperparameters
automl in your continual learning pipeline
20. An idea of AutoML for Computer Vision using Deep Learning
23. • Prepare to manage a whole lot more models.
• Track everything!
• Algorithm, hyperparameters, code version, data version, metrics
automl in your continual learning pipeline
24. • ML Ops
• Automate your machine learning infrastructure and experimentation using
cloud services and Kubernetes!
automl in your continual learning pipeline
25. • Deploy automatically, but carefully
• Run tests before/during/after deployment and define your benchmarks well!
• Test on past (but relevant) data, performance, system
• Use Canary release technique!
A technique to reduce the risk of introducing a new software version in
production by slowly rolling out the change to a small subset of users before
rolling it out to the entire infrastructure and making it available to everybody.
automatic & safe model updates and deployments
https://martinfowler.com/bliki/CanaryRelease.html
26. automatic & safe model updates and deployments
• Use Kubernetes
• Use Istio
• Read our guide:
https://cnvrg.io/deploy-models-with-kubernetes/
27. • Monitor your input data
• Search for unexpected values
• Measure correlation of production data to train data
• Keep it open to dynamically add new tests and rules
• Monitor your predictions
• Model confidence, model bias, and more
• Use Kubernetes, Prometheus and AlertManager: https://prometheus.io/docs/alerting/alertmanager/
monitoring
28. • Your ML pipeline should be triggered based on
a) Periodically (once a day/week/etc)
b) New training data that is coming in
c) Model decay / model bias / alerts in production
• Make sure to track and validate triggers
trigger retraining
30. summary
• Continual learning is the future of your ML models
• Continual learning is doable and worth the investment
• Add tests in every stage: data -> train AutoML -> test -> deploy
• Building it on your own? Connect science to engineering
• “Continual Learning in Practice” - By Tom Diethe, Tom Borchert, Eno
Thereska, Borja de Balle Pigem, Neil Lawrence
• Next: Active Learning, Human-in-the-loop, and more