Machine learning, what’s missing?

David Baumgartner

April 22, 2020

What is wrong?

Is machine learning broken by design?!?

Money Money Money … deployment?

… think about tomorrow …

My encounters

#1: Process

Software Development Process

Machine Learning Process

IBM - Big Data University
IBM - Big Data University
Microsoft
Microsoft

#2: Conventions / Guidelines

Software Development Conventions

  • Git(Hub/Lab)/BitBucket/…
  • Intellij/Visual Studio/…
  • PyPI/Ubuntu Repo/AppStore/…

Machine Learning Conventions ?


  • Target question?
  • Answer?

Tools?

#3 Review & Merge

#4 Tracking

  • Code = Yes

  • Tests = Yes

  • Model = Jain

Result Tracking

  1. Copy/Paste from terminal to a file
  2. Updated to automatic file appending
  3. Log full training to a file

Parameters Tracking

  1. Resulting parameters from final model
  2. …
  3. hmmmm…

Data usage

  1. My data haven’t change
  2. …
  3. hmmmm…

Soups & unfinished

  • Many eco systems

  • Manage and Track


  • Storage

#5 Ship it

Software Deployment


  • Jenkins/GoCD/Drone/Sonar/…
  • Grafana/ELK/…

Model Deployment


  • Custom Code
  • Pipelines AirFlow/KubFlow/…

#6 Export & Import

  • REST deploy (OK)
  • Client/Edge deploy (!?!)
  • GDPR ?!?

Summary

§1

  • Money makes life easier

  • $ == true: go into a cloud or buy our own

  • $ == false: your are stuck to a lot of manual work

§2

  • No gold standard neither for tools nor for a process

  • Very fast growing but no conventions and no guidelines

Thanks and happy modeling