Challenges faced with
machine learning in practice
12:15 Introduction
12:20 Machine learning lifecycle
12:30 Different aspects
12:50 Cubonacci
12:55 Key takeaways
Introduction
Jan van der Vegt
Data science background
Founded Cubonacci in 2018
MLOps platform
https://www.linkedin.com/in/jan-van-der-vegt/
80% of projects intended for production don’t make it
Why?
Machine learning lifecycle
Machine learning lifecycle
Infra
Versions
SecurityLive data
Monitoring
Machine learning lifecycle
Hidden technical debt in machine learning systems
https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf
Machine learning lifecycle
80% of projects intended for production don’t make it
Looking at the full machine learning lifecycle, from
development to maintenance is essential for success
Machine Learning Operations
MLOps goals
Efficiency
Value faster
Collaboration
DRY
Trust
Increased usage
Peace of mind
Process
Aspects
Development
Deployment
Infrastructure
Monitoring
Automation
Standardization
Lineage
Reproducibility
Overview
Development
Iterative
Fast
Messy
Rigid
Slow
Reliable
Deployment
Types:
Request based
Scheduled batch
Streaming
Important:
Frequent
Fast
Autonomous
Infrastructure
Dynamic Resource
based
Monitoring
Feedback
Model metrics
Data quality
Input data
Predictions
Availability
System on?
Requests
Alerts
Severity
Responsibilities
Diagnostics
Logs
Inputs
Automation
Iteration Fewer mistakes Change
Standardization
Automation Collaboration Flexibility
Lineage
Source data
Code
Data snapshot
Experiment
Model Prediction
New data
Feedback
Monitoring
Prediction
Reproducibility
ExperimentData
Reconstructable
Save snapshot itself
Random number generators
Hyperparameters
Input data
Trust
Diagnosis
Regulations
Overview
End-to-end system
Unified view on the system
Communication significantly easier
Trace back any issues to the source
Cubonacci
Flexibility
Design Efficiency
Takeaways
- Keep the full machine learning lifecycle in mind
- Figure out what is important
- Doing nothing fails to scale
- Doing everything stifles progress
www.cubonacci.com +31 (0) 20 308 43 90 info@cubonacci.com

Jan van der Vegt. Challenges faced with machine learning in practice