際際滷

際際滷Share a Scribd company logo

When a machine learning researcher and a software engineer walk into a bar

Paul Lam
Paul Lam

This document discusses challenges that arise when machine learning researchers collaborate with software engineers. It presents three key problems: 1) Transitioning research code to production-ready code, 2) Interfacing machine learning models with software, and 3) Differences in terminology between machine learning and software engineering. The document explores options for refactoring code to be shared between teams, proposes designing machine learning as an internal API, and emphasizes the importance of pair programming to overcome differences in backgrounds and jargon.

1 of 22
Download to read offline
When a machine learning researcher and a software engineer walk into a bar Presented at Data Science Slovenia May 2016 by Paul Lam (@quantisan)
Multi-disciplinary to Inter-disciplinary Data Scientist Engineers Designers Product Managers Users Analysts Ops Project Managers User Experience Marketing Sales Marilyn Stember, Advancing the social sciences through the interdisciplinary enterprise, The Social Science Journal, Volume 28, Issue 1, pages 1-14, 1991
Real engagement makes people happy, fulfilled, heard, and motivated
Researcher, Engineer, Designer, and Product Manager Machine Learning System
Problem 1: Research Code to Production Code Machine Learning System
Research Code to Production Code - Our Options Use a framework, e.g. TensorFlow Pros: Leverage community engagement Cons: Too early for us to be tied to a framework Re-write to Java/Scala/Clojure Pros: Performance, engineering team experience Cons: Performance not an issue for now, double the work Refactor Pros: Least amount of immediate efforts for both researcher and engineer Cons: Best of both worlds or worst of both worlds?
Who calls our machine learning component Machine Learning SystemHere and here
Code sharing with a ML Python module Jupyter Notebook $ pip install burt import burt b = burt.Burt(coll) b.getExperimentBatch() ... Production Server $ pip install burt import burt b = burt.Burt(coll) b.getExperimentBatch() Private PyPI on S3 - https://github.com/novemberfiveco/s3pypi
Engineering practices to consider when refactoring Libraries vs Services Data persistence States vs functions Single responsibility principle Dont repeat yourself (DRY) Documentation Continuous integration Style standards Testing
Share best practices
Problem 2: Interfacing with the ML module Machine Learning System
Machine learning as an internal API RESTful API Machine Learning System ClientService Microservice - http://martinfowler.com/articles/microservices.html
Sequence Diagram - activate campaign System ML External Activate email campaign loadContacts() return contacts initializeCampaign() return campaign ID getExperimentBatch() return experiment Success
API definition: getExperimentBatch() Input Output
Our internal machine learning API [1] swagger.io, [2] flask-restplus
Specifications over documentation
Problem 3: We are used to our own jargons Machine Learning System
Mysteriously failing unit test on getExperimentBatch()
When GET is more than GET
Pair Programming
Pair up
Summary Machine Learning System

More Related Content

When a machine learning researcher and a software engineer walk into a bar

  • 1. When a machine learning researcher and a software engineer walk into a bar Presented at Data Science Slovenia May 2016 by Paul Lam (@quantisan)
  • 2. Multi-disciplinary to Inter-disciplinary Data Scientist Engineers Designers Product Managers Users Analysts Ops Project Managers User Experience Marketing Sales Marilyn Stember, Advancing the social sciences through the interdisciplinary enterprise, The Social Science Journal, Volume 28, Issue 1, pages 1-14, 1991
  • 3. Real engagement makes people happy, fulfilled, heard, and motivated
  • 4. Researcher, Engineer, Designer, and Product Manager Machine Learning System
  • 5. Problem 1: Research Code to Production Code Machine Learning System
  • 6. Research Code to Production Code - Our Options Use a framework, e.g. TensorFlow Pros: Leverage community engagement Cons: Too early for us to be tied to a framework Re-write to Java/Scala/Clojure Pros: Performance, engineering team experience Cons: Performance not an issue for now, double the work Refactor Pros: Least amount of immediate efforts for both researcher and engineer Cons: Best of both worlds or worst of both worlds?
  • 7. Who calls our machine learning component Machine Learning SystemHere and here
  • 8. Code sharing with a ML Python module Jupyter Notebook $ pip install burt import burt b = burt.Burt(coll) b.getExperimentBatch() ... Production Server $ pip install burt import burt b = burt.Burt(coll) b.getExperimentBatch() Private PyPI on S3 - https://github.com/novemberfiveco/s3pypi
  • 9. Engineering practices to consider when refactoring Libraries vs Services Data persistence States vs functions Single responsibility principle Dont repeat yourself (DRY) Documentation Continuous integration Style standards Testing
  • 11. Problem 2: Interfacing with the ML module Machine Learning System
  • 12. Machine learning as an internal API RESTful API Machine Learning System ClientService Microservice - http://martinfowler.com/articles/microservices.html
  • 13. Sequence Diagram - activate campaign System ML External Activate email campaign loadContacts() return contacts initializeCampaign() return campaign ID getExperimentBatch() return experiment Success
  • 15. Our internal machine learning API [1] swagger.io, [2] flask-restplus
  • 17. Problem 3: We are used to our own jargons Machine Learning System
  • 18. Mysteriously failing unit test on getExperimentBatch()
  • 19. When GET is more than GET