“Only 22% of companies using machine learning have successfully deployed a model”.
says a Deeplearning(dot)ai report
What makes it so hard? And what do we need to do to improve the situation?
A thread –
#1 ML is code using data. Not just code.
#2 Output of the models are dependent on the data it could receive during prediction – a behavior that you cannot completely predict
#3 Real world data doesn’t come cleaned in a file or table
#4 Real world problems involve identifying sources, bringing them together and *trying* to get them to talk to each other
#5 So while code development is controlled, data cannot
#6 What is the outcome : longer deployment times, breakdowns, and impact to performance
#7 Enter MLOps – MLOps is the intersection of Machine Learning, Data Engineering and Dev-ops (borrowed from a product SDLC ).
The aim is to deploy and maintain ML systems in production reliably
#8 A successful team ideally should include a datascientist, a Dev-Ops engineer and a Data Engineer. And most importantly work in tandem
#9 A data scientist alone cannot achieve all the goals of MLOps
#10 Why should this matter to organizations – hire for complementary skills.
MLOps allows companies to easily deploy, monitor, and update models in production, paving the way to AI with ROI.
Analytics is not magic.
Garbage in. Garbage out.
Sources and other resources :
https://www.datarobot.com/lp/why-you-need-mlops/
https://towardsdatascience.com/ml-ops-machine-learning-as-an-engineering-discipline-b86ca4874a3f
**************************************************
#Analytics
#MLOps101
#reviewswithranjani