The rush to Machine Learning for a startup.
Artificial Intelligence, Machine Learning, Neural Networks - what ever you want to call it - it still seems to be the thing de résistance for startups at the moment. Granted, a little less so as we are deeper and deeper into the mess that is Covid19, but a thing none-the-less.
I love what Machine Learning offers, particularly startups, and particularly now that it is becoming super accessible. There are great tutorials, Google even has a crash course! (gently directing you to their products of course)
My concern is that a lot of the time, an algorithm suits just fine to validate your thinking or product. And I (yes, anecdotally) have seen more than a few startups that I have spoken to drive themselves into the ground because of the "Machine Learning or bust" mentality.
Building a suitable Machine Learning based solution requires a solid understanding of what you are trying to achieve, with the tools and the data you have at hand.
It is a maze of dead ends and false starts, and without a solid understanding of what is genuinely possible with data - sometimes it's too hard to even know that what you are seeing is working (or not).
Pioneera is an Australian company working very hard to reduce the impact of stress in the workplace. One of the early approaches here was to use algorithms processing the type of language being seen and presenting that data, rather than embarking on what could easily have been a year long, zero revenue generating, machine learning project - with no assurance of success.
Apart from getting a product to market faster, there have been a few big benefits for Pioneera.
The hypothesis that language and wellness in the workplace are linked has been shown to be anecdotally true. Customers are reporting that the visualisations produced in Pioneera's dashboard correlate with stress and wellness events in their team.
As well as early validation (and thus early revenue) starting with an algorithmic approach has allowed the Pioneera team to better understand how Machine Learning can be applied in the future for better outcomes, and how to test the validity of that application.
The application of Machine Learning to a project has traditionally imbued a project with an opaqueness that some potential clients or investors will not be able to stomach.
If you can't clearly explain how your product is doing what it says on the box, how will you build trust? Saying "It just does!" doesn't cut it a lot of the time.
Working with your data, understanding it, building reliable processing algorithms and developing code from that which is reliable and reproducible builds a very solid foundation on which to grow.
I'm not saying that algorithms should always be the first port of call when you are starting on a data heavy project. In my experience though, Machine Learning (almost always) is not.
Disclosure: Troy Kelly is the CTO of Pioneera Pty Ltd