In my opinion, a lot of these problems are organizational in nature as the tech industry wrestles with how to integrate data science into product development.
Too often there are siloed "data science teams" who do the best they can to field asks from tons of different business areas. Within those teams there are handoffs to labeling teams run offshore in MTurk style, handoffs to data engineers to build ETL to feed the modeling efforts, and then when models are finally trained there's a handoff to machine learning engineers to implement them!
Clearly data scientists need to be better integrated into product development flows, and not treated like a strange specialized black box.
This idea that only creating models is important work is pervasive. I have seen mathematicians, the director of... or vice-president of..., trying to manage complex software problems just focusing on their models, holding all the company power, and disregarding the most basic engineering practices. That projects fail badly, like the waterfall projects in the 90s.
If you are a mathematician that trust software engineers and work with them to create a successful project. We need more like you. I'm tired of arrogant ones that make what could be the most interesting projects a nightmare.
The article hints at one potential source of the problem when it talks about how this doesn't fly for civil engineers, namely a civil engineer needs to be licensed, but anyone can be a data scientist. When I was doing my master's in stats, we spend tons of time on data sampling, things like Simpson's paradox, and other subtleties. But would expect the same things from someone who graduated from a DS bootcamp?
One of the other issues is that we have very few dual expertise scientists/experts. Someone who is a domain expert in both battery manufacturing and AI, for example, needs to be the person in charge of using AI to predict battery problems. However, domain expertise is difficult and time-consuming to acquire. In a metrics-obsessed organization, the time to develop expertise is considered non-productive, so nonsense propagates.
5 comments
[ 2.7 ms ] story [ 25.9 ms ] threadToo often there are siloed "data science teams" who do the best they can to field asks from tons of different business areas. Within those teams there are handoffs to labeling teams run offshore in MTurk style, handoffs to data engineers to build ETL to feed the modeling efforts, and then when models are finally trained there's a handoff to machine learning engineers to implement them!
Clearly data scientists need to be better integrated into product development flows, and not treated like a strange specialized black box.
If you are a mathematician that trust software engineers and work with them to create a successful project. We need more like you. I'm tired of arrogant ones that make what could be the most interesting projects a nightmare.