This talk is meant to look at the (increasingly know?) lifecycle of a product with Machine Learning at its core, step by step. For each of those steps (of which model training and tracking may be just a tiny part of the work), we will look at the pitfalls that may keep it stuck in prototype phase and suggest open source libraries that can help.
This talk uses a single tweet of frustration against "AI and COVID" as an excuse to look at the presence of Machine Learning in production. Where are we in 2021 in terms of real world solutions? What does the lifecycle of a product with ML at its core look like and, at each step, what issues get it stuck in prototype phase? For each of those steps, are there tools that can get you in a better position to move to Prod?
If we're all tracking experiments now, what's missing?
The response this talk suggests for that last frustrated question is "Everything else, which often matters a lot more than the model itself. But there's tools for that too".
The talk makes quite a few assumptions about the audience: you are likely a developer or researcher (also at a senior level - nobody's clean); you are aware of steps necessary for a Machine Learning project. Trickier yet: you agree that a low product:POC rate is of limited use in the industry, and you would rather use existing tools whenever possible.
We will look at the full lifecycle first, from conception of the solution to modelling to deployment, including typically forgotten steps like incorporating the user's point of view. More importantly, we'll see prototype pitfalls for each step that can get you stuck and, more importantly, tools and libraries that help you to overcome those faster.
This cannot be a complete guide, but at the end of the talk you'll hopefully...