Building a Next-Generation Machine Learning Platform in the Cloud

"Are we asking the right question and drawing on best practices?"

Going back to a discussion with an executive, I was super excited that my customer is building an AI platform. However, the question I got asked was would their teams adopt it? And that brought me to even a deeper question – Does a platform sets you up for success in AI and AI driven-organization? It reminded me of a similar situation a while ago when I was working with one of my favorite CPG clients wherein we built an insights platform for manufacturing facilities. And the same question was asked – would the manufacturing facilities adopt it?

I think we need to draw a parallel line just like how we do it with the Industrial Revolution and AI revolution. We all forget that many of these platforms remain the same and there is a huge adoption hurdle that you need to cross. Self-Serve, Ease of Use, Discoverability, and Change Management are probably the keywords. How do you manage change, how do you change the culture, and how people think becomes crucial to adoption.

Here are some of my learnings on building a few AL/ML platforms

• Amalgamation of Art and Science: Data Science and Machine Learning is an Art and a Science, and we need to make sure our data scientists have the freedom to explore, to fail and then to try again. What else gives a better chance for them to succeed without being stuck except for a platform based in the cloud.
• Simplicity: Make it Easy for data scientists to get access to your enterprise data, commercial data, and cross-client data so they are not spending weeks just trying to access datasets for their experiments. Have you thought about access through data lakes?
• Living at the cutting-edge: Changes are happening fast in AI. Sometimes I have a hard time keeping track of all the cutting-edge innovations and models. Does your team have easy access to these new technologies without much of the sunk cost? For example, AutoML can help them get a better understanding of data and insights into algorithms faster. Perhaps, how about the use of commercially available AI services for example OCR capabilities for various languages.
• Model Sharing, Reuse, and Chaining: Search and reuse mantra, it works. Is your organization sharing models that have been developed for one use case but could be leveraged for another, perhaps chain them together? Can that be done easily by simply signing up vs spending weeks of development effort in reusing the models? The long-tail ROI of existing models is a hidden treasure.
• Run fast and scale faster. Can your platform scale not only predict faster but also train faster? The more experiments that you can train the better and faster you can run your enterprise on AI and create a competitive advantage for your clients.
• Remove the barriers to Deploy and Retraining: It is easy for data scientists to deploy a model and perhaps retrain the existing model. Does your platform solve the ML operations hurdles?
• Fair Learning or Black box: Is your platform providing you insights into bias, harm of allocation by preferring a certain group, or harm of quality of service by working well for a certain group within your customer base. Can your platform accommodate disparity metrics and parity constraints? Can we go back and easily Audit for Fairness?
• Workspaces vs VM: Traditionally it was easier to just give a GPU or a high-end VM to a data scientist, however, modern platforms have come far ahead. Separation of workspaces and computers to run your model is not only cost-effective, but also lets sharing and learning of the models easier. You only pay for computing that you need to run your training and models. Most workspaces are available free on platforms.
• Self Service: It is the way to go to enable teams and data scientists to quickly access the model without weeks of delay in either helping to test or deploy. Are you building systems where teams can automatically subscribe to a build ML model or its API? Do we have the right documentation to drive adoption without reaching out to the human in the loop?
• CCT: Build mindshare through Community, Champions, and Training. A community of data scientists through workshops is very important in influencing and peer to peer learning. Continuous training helps us keep abreast of new developments and technologies that could be easily adopted.

Author: Nupur Srivastava

Microsoft: Director – Machine Learning and Artificial Intelligence Azure Cloud