The Catastrophe of ML Success

It may be unclear that your ML path is a dead end. Cape Palliser, New Zealand

So you’re exploring a gnarly problem using MLaaS (machine learning as a service), and you create a model of amazing accuracy and precision. It will be huge for strategic advantage. This is fantastic, way outside expectations.

What next? After you tell your colleagues and the buzz spreads through the company, everyone wants the model in their apps. Some apps require batch prediction feeds, others are real time. The data is huge, possibly bigger than Big, and so proprietary and strategic even the database names are Top Secret.

And this is the potentially catastrophic moment. Everyone is wild to get something you might not be able to give them. You started out with hypotheses and experiments and explorations. You found a ML service with functions that support your work, but never asked how to take your work home with you. You didn’t anticipate success, so you didn’t ask:

   How do I export this model to my enterprise, so it can be operationalized? 

   How do I replicate the model on other platforms, that are closer to my data, or more private, or in my control?

Bottom line: Production use of your model drives critical requirements for selecting a machine learning service. Frustrating: really, you just want to get to work with your experiments. But the pain of being stuck in a cul de sac of your own making is far greater.