Let’s face it.
The current media headlines have all managed to make us believe AI is the Alohomora to all your problems: AI will open all doors that were previously locked and unpenetratrable by your baseline systems in use. From detecting early cancer signs, to accelerating drug discovery, to optimizing revenue and sales streams… all the way down to what else would optimize your metabolism on top of your morning nutella spread.
Just like similarly, media managed to make us believe Bitcoin is the door to richness, putting forward the arbitrage opportunity and failing to dive deeper into the complexity of the underlying blockchain technology.
Same story with startup hype in 2015 and beyond, which led to a ridiculous amount of birthed startups that don’t really solve a problem (but instead look for problems to problems, or problems to solutions), fail to understand users or even build enough traction before they go bust (not to say these failures will not teach a valuable lesson, but some of the “silly” failures could have been prevented by a stronger, more proven methodologies. Most successes you’ll hear about are only examples of survival bias).
As a result, only the talented, seasoned, real and down-to-earth practitioners manage to survive, because they understand the underlying technology or complexity of the task, and the necessity of long-term persistence and efforts to have a slight chance of reasonable and satisfactory results.
These people survive, long after the hype has faded away. Be it entrepreneurs, investors, managers.
AI and ML is no exception to that.
No one will contest these technologies are real and possess tremendous impact potential. But the truth is: most companies aren’t ready for AI/ML. And most managers with little to no experience in data science will be called on to make the hard decisions, internally and externally.
What managers need to do is to refrain from getting stuck in the hype pool and needing to include an ML aspect in their workstreams or business value no matter what, to keep up with market hype and peer business pressure — especially when they don’t need it.
So let’s see how they do.
Issue 1 — A Problem Problem: The type of problems that are being thrown at ML are not technically and realistically appropriate and thus management has little prior understanding of what is essaentially solvable, what is feasible and what is not.
Managers are not to blame. Who doesn’t want to set ambitious goals for the company? But what if these ambitions are set on unsolvable problems? The ambition bitterly turns into disappointment.
Unsolvable problems are merely unsolvable due to the underlying nature of irrational markets or human behaviors.
Let’s take an example for Stock Market. or heck. Even housing prices. Suppose you run the model in 2007–2008. Your model will be full of green indicators based on a run-up in housing prices fueled by demand, speculation and lavish spending.
Then the bubble bursts. But your ML algorithm couldn’t know any better because it learned only on historical data, and it could not predict irrational markets or unpredictible policies.
Basically ML will be good at predictions (namely classification and regressions), but is really good at those when it comes to well thought out and defined problems, given a near-complete set of feature space that is directly correlated to the outcome sought.
Companies fall into the trap of traditionally collecting the same data over the years and expecting different results to different problems. They want to look at the data and figure out the problem to solve. Or start with the data and force it onto existing problems, disregarding the necessity for different approaches to solve the latter. Two very wrong (and expensive) approaches to problems. Well, Einstein has a word for you here:
“The definition of insanity is doing the same thing over and over again, but expecting different results.” — Einstein
Issue 2 — A People Problem: Failure to trust and empower the data science team to build a guidance for what can solved and areas of improvement where their expertise is relevant.
Where less data-driven management tends to fail is considering their ML/Data teams as a sort of black box lab where they send their untreatable problems or would-like-to-haves to be figured out. No questions asked, but a lot of ambitious expectations upfront.
In the end, the people who suffer most are data scientists who come in with great expectations to help and support business units in their companies but are faced with a wall of unreasonable expectations, and worse, a lack of understanding of the complexity of the tasks undertaken to solve unsolvable problems. It is only a matter of time when managers would think they hired one too many data scientists and that the business in fact, did not need that at all.
Issue 3 — A Data Problem: the industry or the company as a whole is just not ready to move forward to ML.
The ML and Data Science on Kaggle is totally different ball game than reality. You will get nowhere near exact problematics like these, with readily available (and cristal clean) datasets. Unless we’re talking about proven problems that are repeatable and have clearly and traditionally showed a correlation with a set of existing features, there is little evidence that whatever thrown features out there will help build a reliable algorithm.
So if your feature space is not correlated with your target variable (meaning the goal you want to optimize for), then you are in bad luck. You’ll end up predicting your nephew’s company turnover with the number of bananas your girlfriend had for supper last year.
Hence why it is important to pick the right set of features, internally and externally, through the company’s own collected data and readily available open data…. while being smart about it.
Open data is a great leap forward, but trying to solve a prolem with common census statistics alone is not enough to predict future profits/asset appreciation or depreciation, simply because there are a number of hidden features that may or may not be collected, may or may not be initially even thought of to account for a percentage correlation in the outcomes.
Advice to Management — Your future AI-Hype-Proof Business Relies on Three Pillars:
Problem, People & Data
Bottom line is, the key to success is first, the ability to translate a business problem into a technical one, and second the ability to translate the adopted technical procedure to a business solution.
This means, the ability to define the right problems, with the right people, with the right data.
What managers should rather do is focus on proven problems that have clear ML solutions given the right set of feature space (provided the company can provide those datasets). These problems tend to be recurrent problems in any businesses and many automated solutions have already been developped by the big companies cloud solutions (Amazon, Google, and Microsoft). Instead of reinventing the wheel, integrate these tools into your work streams and have data science/data engineering teams maintain the process while devoting a percentage of their time to figuring out, and discovering innovative processes that are akin to the business and industry. Essentially, consider the data science endeavor as both a maintenance & data-driven business support unit while fulfilling its R&D role.
Finally, leave a margin for failure, because the quest to AI is a discovery quest, whether discovered methods may or may not reach a dead end.
First, you learn what doesn’t work.
Second, you learn what may work elsewhere.
If you found this article useful would like to learn more, I’ll be posting a new article shortly about addressing the three pillars above- so stay tuned or follow my medium channel.
I’d appreciate you share the article with someone you believe needs to know about this. Be it a manager or a data scientist who can share this with their team, or just an AI/ML enthusiast who refuses to buy into the hype.
Thank you for reading!