Operationalizing Machine Learning — Decoupling Engineering from AI

Previously I wrote about how today’s data science tooling has changed. I recently reviewed several vendor tools all claiming to enable “faster iteration and deployment” for data science projects. There are many more out there that I haven’t seen in much detail, but the general trend is promising:

“Empower data scientists to control the end-to-end AI lifecycle by decoupling engineering from AI.”

But why? (We’ll cover “how” in a later post!)

To answer the question, a group of Berkeley researchers recently published a study (2022) that summarized the opportunities and challenges of machine learning operations across industries through semi-structured interviews. As shown in the below table, the 19 participants of the study represent small to large companies spanning across various industries including autonomous vehicle, computer hardware, retail, Ads/marketing, finance/fintech, bioinformatics, cloud computing and deep learning. I also attended a discussion about this study hosted by Hamel Husain and Josh Wills with one of the researchers — Shreya Shankar.

The motivation of the study was to seek “clarity to MLOps” and “a richer understanding of best practices and challenges in MLOps” that can…