My previous blog post considered what we can learn about learning by watching computers master video games. What’s interesting is that the computer in that case wasn’t an “expert” at video games. It hadn’t made a comprehensive study of all the different kinds of games and ways they function – it conquered the game by playing.
In other words, subject-matter knowledge isn’t really so essential to solving certain kinds of problems, even highly complex ones. What’s essential is a process for problem-solving. A computer that can visually detect cancer cells isn’t “recognizing” sick cells: it’s telling the difference between a cancer cell and a healthy one. It’s practicing pattern recognition.
One of our engineers spent a weekend putting together an intelligent app that could classify cereal brands. Point your phone at a cereal, and the AI will tell you what you’re eating. It comes out like a weather report: 79 percent chance of Sugar Crisps, 10 percent chance of Cocoa Puffs. (If you want to give it a try, it’s available in the Google Play store.)
A single AI can have many applications
A fascinating takeaway of all this is that because specific domain knowledge is not involved, you could ‘retrain’ a cereal classifier to instead tell the difference between kinds and sizes of nuts and bolts, or pieces of cutlery, or fruits and vegetables. A machine designed to do one thing may be capable of a half-dozen others.
Now, interestingly, where domain expertise becomes important is in evaluating or incorporating results. Domain expertise or subject-matter knowledge can help determine if any bias has found its way into the model, which is possible. For example, if the cereal classifier had been trained using data that’s 99 percent cocoa puffs and 1 percent every other cereal, the results would skew. Also, if a basic version of this cereal classifier was pointed at something that was not cereal at all, then it would still assume it was identifying a cereal and come up with a prediction within a range of cereals, like Cocoa Puffs or Sugar Crisp. Somebody needs the knowledge to be able to say, “Yes, that is cereal” or “No, that’s not cereal.”
Pattern recognition can help keep people safe
With a bit of tweaking, the cereal classifier could be trained to do something really useful, like recognizing cancer from patient tissue slides. Human expertise in this area is varied, skilled pathologists can disagree a lot of the time. When trained on thousands of slides, a model similar to the cereal classifier with much greater accuracy than human epidemiologists. The classifier isn’t a medical expert, nor is it memorizing or using a rules-based approach. Instead, it synthesizes information out of thousands of features simultaneously and learns to recognize patterns.
What does this mean for domain experts? While ultimately many jobs could become automated, the real power of AI is to synthesize its results with human expertise, so that human expertise can be put to better, more productive use.
– 607 words