However, as Peter Xenopoulos pointed out:
Machine learning … Everybody talks about it. Only some really know how to do it. Everyone thinks everyone else is doing it. So, everyone claims they’re doing it.
As many have painfully discovered simply hiring a data scientist and trusting something good will happen is an expensive form of window dressing that doesn’t deliver real business value.
Here are ten steps to help you down the path from “claims your doing it” to “really knows how to do it”.
Asking the right question
As Clayton Christensen reminds us “asking the right questions is more important than getting the right answers”. There is no place this is more true than a machine learning project.
Artisanal vs. industrial
When solving a class of problem that has never been solved before an artisanal hand-crafted approach makes sense.
When building an organization that can efficiently and repeatedly solve the same class of problem across a variety of specific circumstances then a more industrial approach is warranted.
Many hurdles to scaling up machine learning come from not recognizing when to transition from artisanal to industrial.
Talking data science
Stakeholders are intimidated by data science and data scientists enjoy the mystique of an inscrutable craft – this is not healthy.
Stakeholders need to learn sufficient data science concepts and terminology to break into the silo. Data scientists need to recognize that effective communication is just as important as a well-tuned model.
Data science surgical teams
Surgery is a high stakes activity where carefully honed skills have a huge impact on a highly variable outcome. There are good reasons why surgical teams have well defined specialized roles.
These lessons apply to machine learning projects. Segmenting, professionalizing and optimizing the different aspects of data science projects will help us turn machine learning into a team sport and get past the idea that we can’t scale up because data scientists spend all their time as janitors.
More than a model
Often machine learning projects focus too much on the middle act, defining and optimizing the model, For a project to live up to its potential we can’t treat everything else as an afterthought. We have to succeed at data gathering, feature engineering, validity testing, production deployment, etc. and think holistically about how all of these elements come together.
Emergent behavior from feedback loops
Perhaps biologists and game designers have more in common than we assume. For example those two professions have thought the hardest about “emergent behavior” and they agree that getting rich positive results from complex multi-actor systems requires having the right feedback loops in place.
This insight applies to machine learning initiatives: it is all about the feedback loops. What you learn in each phase of your project can help you refine and optimize previous phases. It is critical to have an agile approach that can treat insights from feedback as assets to be incorporated rather than annoyances to be minimized.
Treat future you right
It is rare for a high value machine learning project to be a one-time event. We want to pass the marshmallow test and recognize the value of deferred gratification. We need to think about continuous improvement in a changing environment where we demonstrate increasing efficiency while delivering a stream of related results.
Better data beats fancier algos
Of course it is valuable to chose the most appropriate algorithm and tune its parameters optimily. What is even more valuable is getting better data to feed that model.
Insight outweighs predictive power
Of course it is valuable to create a model that demonstrates high predictive power on your current data set. What is even more valuable though, is generating new insights into the underlying system you are modeling and using those insights to generalize and optimize your model.
Explainability is table stakes
Today Explainable AI (XAI) is treated as a specialized sub-domain. We have that arsy-versy. More and more we will see explainability being a central requirement, with the use cases that can be black boxes being the exceptions.
P.S. This post was going to be titled “What we talk about when we talk about machine learning” until Gawker slapped some sense into me.