Machine Learning can’t learn the future

Charlie Hubbard
Nov 6 · 4 min read

Machine learning is about making predictions about uncertain things, and we think this is a wonderful thing. However, how a machine learning algorithm works, at this moment, it must learn from historical events. That past influences its future predictions. While this is a form of intelligence it is not the only skill of intelligence. It isn’t even how humans learn new skills either.

My son, as most boys do, loved fire trucks. We would drive around and I’d point out firetrucks I saw. He got so interested and excited to see a firetruck, so it was a lot of fun for me as well. Eventually he began pointing them out, and asking me to take him to places where we’d seen them before. But, while this occupied our time twice a day for close to a year, he never saw more than 100 examples of fire trucks. And probably less than a dozen before he was able to pick out firetrucks. His ability to pick out fire trucks was generally robust. The environment did not matter he could complete the task without much error regardless if it was raining, cloudy, or full sun. I remember only one time he pointed to a red truck and said fire truck, but it may have been because we were talking about the difference between firetrucks and the fire chief truck. So he may have said firetruck meaning fire chief truck which is a forgivable mistake.

My son’s ability to recognize fire trucks with a high degree of accuracy and recall isn’t unique. This is not some proof of a gifted child. All children can do this reasonably well. And they never have to see 1000’s of examples before they can pick out firetrucks or whatever with even moderate accuracy. This simple comparison shows that how humans perform pattern recognition is very different than how computers perform that task.

Computers have to have large amounts of examples before it can get even into 70% recognition. All of these algorithms existed going back to the 1950s, but what changed was our ability to amass large quantities of examples to feed to these algorithms in recent times. This suggests that large quantities of data were required to make these algorithms even moderately useful. Even with lots of data tasks like speech recognition, OCR, and hand writing recognition are still just passable using todays latest technology.

If you only use history as your model for what is a valid prediction. You can’t imagine a prediction that you never seen before. Machine learning algorithms only recognize patterns, but only the patterns it has seen before. It possesses no ability to understand why the prediction exists or why they are plausible. It just knows that it has happened before therefore, it must be plausible. As usual computers do not possess or model why. What underlies all of this is what is likely to happen given what’s happened before, but never to predict another state that hasn’t happened before.

This ability to conceive of something unseen that humans possess is not understood well enough to model as algorithms. Generating new possible futures that could be answers to a prediction is harder than simple pattern recognition. And while I place pattern recognition at a somewhat lower level than generating new outcomes, machine learning algorithms are superior to humans ability to recognize patterns. Recent discoveries of new scientific value were made purely by training a machine learning algorithm on abstract texts of thousands of research papers. The computer had no prior knowledge of material sciences or any other scientific principle, but it was able to find a pattern in the scientific literature that humans had not recognized. That is an amazing discovery, and demonstrates machine learning’s superior pattern recognition performance to human’s. But, similar to how computers are superior at arithmetic to humans that doesn’t make humans subjugated to our superior arithmetic overlords. Humans still needed to produce the scientific papers to be the input to the awesome power of pattern recognition that computers could perform.

The ability to scale pattern recognition on a mass scale to cater to individualized tastes by putting people into predictable groups is called collaborative filtering. While that process is efficient, it is fundamentally a conservative way to produce predictions. In that, the system thinks if you liked this then you’ll like that because it’s seen this pattern before in this group. The system won’t choose to show you something new and different because it can’t imagine what that is without risking it’s accuracy. It lacks the ability to predict a future where you’ll enjoy its prediction that the group hasn’t seen. And without that ability it just simply keeps giving you what you like, but it’s just the same thing over and over.

The way to overcome this conservative view of predictions is to retrain your machine learning algorithms periodically. Retraining models hopes that the system can evolve forward by taking into account new data unseen when it trained last time. This can be a middle ground between the always recommending the past vs. looking to the future, but it still requires humans to discover the future to provide to the machine learning training process. Without human changes in taste or interests it can’t change.

In some ways this should be comforting about our future with machine learning. It still needs us. However, this static nature of these algorithms is also quite dangerous in how it shapes us culturally because poor implementations of these algorithms risk preventing progress and evolution. How can that usher in a brilliant new future if the future is amazing it’s just like what you remember? This ultimately why machine learning most likely will not usher in some amazing new future with better options for humans like so many artificial intelligence hype men sell us. It’s not that artificial intelligence is evil, but that it is limited in serious ways that humans do not have.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade