Dear fellow C-person,
To save you time, here’s the summary of this blog post. You know very well that there is a golden “middle path” between the two extremes of losing-the-focus and staying-the-course. In machine learning this is known for decades as “exploration-exploitation tradeoff”. (Scientific terminology is not PC, but at least it is consistent). This blog explains what you can learn from machine learning, and why you should listen to your scientists and engineers.
An old story plays over and over again: a company pursues a short-term strategy of cost-cutting, creative book-keeping, or “staying the course”, instead of investing into strategic goals, core employees, research, and infrastructure. And then, almost invariably, this greedy myopic vision costs that company dearly. Consider a rodent: a rodent smells the cheese, a rodent finds the cheese, a rodent takes a bite, blam! goes the trap. Rest in peace, poor rodent. Second rodent smells the cheese, second rodent finds the cheese, second rodent takes a bite, meow! goes the cat. Rest in peace, poor rodent. The moral of the story: think ahead.
Now in more detail
Thinking ahead is actually very, very hard. It is one of the key problems both in real life and in Artificial Intelligence. Let’s look back at the mouse: the mouse would be best off if it could find some cheese with no traps and no cats. The problem is, this requires exploration, and exploration is inherently both dangerous and time consuming. The way this is handled in machine learning (reinforcement learning, to be specific), is that the behavior (“policy”) tries to maximize not the immediate reward (cheese!), but a value function such as having a supply of food for the family. Value function takes into account not just an immediate reward, but also the predicted future rewards (both positive and negative) that follow from that decision you’re about to make. Greedy algorithms only look at the immediate reward. This ends badly except in the simplest or most clear-cut cases. Non-greedy algorithms try to look ahead. But here’s a catch: (a) the future is inherently unpredictable, and (b) you can never explore all the possibilities. There are many ways to consider, yet not to overthink, the implications of any action, but essentially what happens is that longer-range predictions are discounted more, according to some rule. This is a sketch of reinforcement learning in a really, really tiny nutshell.
Adopting AI in general, and machine learning more specifically, as a part of the strategic toolkit of your company will not give you an immediate reward. Rather, it will give you tools to address your pain-points and improve your KPIs based on the data you already have, or on the data you should be collecting. And ultimately it will make your company much more valuable and much more competitive. There are no immediate solutions in AI, and no magical “products” or “platforms” that would answer any questions asked if you rub your wallet the right way. There is, however, a legion of so-called “expert” consultants who are delighted to sell you some woo pseudoscience and black-box AI that would start spewing nonsense as soon as it encounters new data. That’s not the right way to solve your problems or to make your company more valuable. The right way – and this may sound counterintuitive to you at first – is to talk to your scientists and engineers. Skip the middle management and go talk – and more so listen – to your engineers. They will tell you exactly where the pain-points are, and which ones of them can be addressed with machine learning. Get a solid idea of what is doable, what is the potential scope of the effort, and what is its potential value for the company. Then you can make educated decisions about how to implement the ML / AI effort for your company, what the focus and the goals should be, and what to expect. We can help.
About the Author
Dimitry Fisher is a Chief AI Officer at Analytics Ventures / Dynam AI. Dimitry has 20+ years of R&D experience in academia, government, and private sectors, spanning multiple areas of physics, neuroscience, ML and AI. Dimitry received his PhD in Physics from Weizmann Institute in 2002. He did independent research in physics before shifting his focus to neuroscience and to mechanisms of learning in neural networks. Dimitry joined Brain Corporation in 2012, doing bleeding-edge R&D for Qualcomm, DARPA, and several internal projects, including the first self-driving janitorial AI. Dimitry saw Brain Corporation grow from 10 people to 100 before resigning his position to join Analytics Ventures as the Chief AI Officer. He has 12 granted US patents and over 40 publications in scientific journals.