In our AI primer, we differentiated AI and Machine Learning (ML) and gave an example around disease diagnostics. We’ll reference that example here. Because AI and ML are often used interchangeably by marketers, we’ll stick to using ML here to keep it simple.
As with most computer programming techniques, machine learning takes its roots in mathematics. The concepts behind machine learning date back to the first half of the 20th century. As cloud computing capabilities and processing power has grown, what once seemed purely theoretical have become more and more practical. The application of ML has grown steadily since then behind the scenes, with a drastic uptick in the last decade and a surge in press and analysts bringing the term ML mainstream over the past few years. Naturally, marketers took it one step further and started applying the ML label to technology that is very different from the scientific definition discussed in the first article. Over the remainder of this article, we’ll examine some common imposters, how they’re different, and if it really matters.
Automation can leverage ML techniques, but automation in and of itself is not ML. Programs, bots, machines, etc. can be doing the tasks of humans, but this does not make them “intelligent.” They are simply doing exactly what they are programmed to do. The software that underpins blood testing is doing the analysis work instead of a doctor, but it is not intelligent.
You might ask, “Well, what about bots that have to make decisions rather than just following a set routine?” Although this is a step in the right direction, it’s still just a set of “If this then that” rules. If someone configured a bot to do something in a certain situation, it is not “intelligent.” This is similar to the disease diagnosis rules in the first article. If a patient has a cough, congestion, and runny nose, then they have either the flu or a common cold. If they also have a fever, then they have the flu. The conditional rules are static and don’t change.
You might ask about probability- based systems that use complex rules which go beyond conditional rules. This is not only the basis for machine learning, but also, in some cases, identical to how a machine learning model would operate. The key difference here is the “learning” piece of machine learning. Does the system produce better results over time as it gets new information or does it produce consistent results in spite of new information?
Some of you might say “Our system gets updates that improve accuracy. Why is machine learning any better?”
Most importantly, the updates are restricted to improvements that developers are aware of. Maybe they need to add a new conditional rule for an edge case that they discovered. With machine learning, the system is constantly reconsidering its entire internal model. A doctor has a certain base of knowledge that they learned from med school and that they continue to update from interactions with patients. But a diagnostic algorithm has the potential to be better than a doctor because it considers connections that might be meaningless or were previously unknown to common medical knowledge.
Secondly, updates are slow. Companies need to get the data, analyze it, code upgrades, test them, etc. Machine learning systems are limited only by the amount of processing power that they have. They can practically update themselves in real time in reaction to new data.
As with anything else, if you’re paying a premium for it, it should absolutely deliver incremental value. There are solutions out there that are doing amazing things with ML and others that just slapped an ML sticker on for the marketing value. There are solutions that don’t have ML and don’t claim to that are better than their competitors that claim to.
One thing is true without reservation. Systems that leverage true machine learning will perform better than those that don’t when given the same task. It will become the de facto standard over time.
It wouldn’t be fair to end this article without calling out where real machine learning has been successful. It’s not all marketing hype, and where it actually exists, it’s really cool.
Image recognition has reached unprecedented highs. Some photo apps like Google Photos allow you to search your photo library based on what’s in your photos. They’ve also leveraged this image recognition technology as part of their Night Sight camera offering. Rather than improving the hardware on their cell phones (e.g better camera), they used software to artificially lighten photos taken at night. Next time you get asked to “select all of the pictures with fire hydrants”, know that you’re helping improve image recognition.
In the next article, we’ll dig more into the specific challenges that are preventing true machine learning from being leveraged to its fullest potential.