Let your ROI drive a fact-based decision-making process
Use cases abound for machine learning in the enterprise. A lot of the steps and processes are definable and repeatable based on the sheer volume of use cases that have been utilized across many different enterprises.[1] You have to start getting your bearings around how those use cases can be operationalized. During that process is the right time to let return on investment guide you along the way to fact-based decision-making. Getting into that groove is about enabling defined and repeatable processes to deploy machine learning in ways that meet the needs of the organization's machine learning strategy statement. One strategy statement could be as simple as to, “always provide a machine learning driven recommendation for alternate purchases before the payment part of the workflow.” You can begin to work both forward and backward from that strategy. It would be pretty easy to line up a budget item toward making that strategy reality. It is also a well known use case for providing alternatives based on fairly standardized models at this point for recommendation engines.
Getting fact based is about letting the data drive your decision-making process. You can get a list of common use cases for machine learning that exist in the wild right now and are commonly used. Based on those common use cases you should be able to quickly correlate them to your workflows and begin to engage in some return on investment analysis on what it would take to apply machine learning to your workflows. Generally speaking, the start up time to use a machine learning API from GCP, Azure, or AWS is very low.[2] Building out your own infrastructure and getting setup to attach to your workflow with a pipelined machine learning setup to server models would take a little bit longer. You also don’t get the advantage of the scale, tuning, and investment those other companies are making their API setups. However, the general API does not have any customization for your use case if you require something divergent from the nature use case the model they are serving was designed to solve. Some really good models designed to evaluate machine learning cost already exist.[3] Those models are actually being used in the wild of production with more than half of respondents in one survey saying they could demonstrate actual ROI from what they were doing.[4]
Go check out and subscribe to this podcast on MLOps called “Delivery, Interrupted” https://anchor.fm/delivery-interrupted/episodes/Introduction-to-MLOps---Part-1-eqmob5
Footnotes:
[1] Here is a fun video about the common machine learning steps
[2] Building a model from scratch
What’s next for The Lindahl Letter?
Week 6: Understand the ongoing cost and success criteria as part of your ML strategy
Week 7: Plan to grow based on successful ROI
Week 8: Is the ML we need everywhere now?
Week 9: Valuing ML use cases based on scale
Week 10: Model extensibility for few shot GPT-2
I’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed reading this content, then please take a moment and share it with a friend.