In our first blog post in this series we looked at the difference between machine memorization and machine learning. So what does that learning look like in action, and what are the implications for data science?
There was a great example a few years ago, when an AI system was loaded with 49 different Atari 2600 games. The machine was given the games only, not any instructions for how to play them. It mastered every one of them, eventually outperforming humans in many cases. How it did so gets technical pretty fast, involving deep artificial neural networks, reinforcement learning and a few other techniques. But what matters most is that the AI taught itself to look for patterns when gaming, turned those patterns into strategies, and that these strategies evolved over time.
Trial and error
When most humans learn to play a game by concentrating first on the rules. In one video game, for example, one button makes you jump, another fires your blaster. If the bad guys touch you, you lose a life.
But ultimately, just following simple formulas with these rules won’t get you very far. You need to learn patterns: what to expect, when to jump, what to blast. And most gamers need to lose more than a few times before getting the hang of it.
This is true even for a simple game like Breakout. In Breakout, there is a paddle at the bottom of the screen and a bunch of bricks across the top. A ball bounces around, and the goal is to make sure it bounces off the paddle at the right angle to break the maximum number of bricks.
Now, imagine trying to play that game without knowing what a ball is. Or a brick. You know the controls exist, but have no context for what they do. That’s how the AI started. All it knew was that the goal was to get the highest score.
Not surprisingly, its first 10 minutes of play were terrible. But two hours later, it was an expert. Two hours after that, it started strategizing: determining, for example, that it could use the ball to punch a hole in one side of the brick ceiling, angle the ball through that gap, and get it to bounce between the bricks and the top of the screen, scoring points without having to do any extra work.
The machine mastered the game in 240 minutes.
Game learning applied
This type of learning has been used for more consequential purposes as well. A team at Stanford University, for instance, gave a computer 2,000 cancer detection scans to look at. When humans do this screening, their eyes search for light and dark spots. Skilled pathologists can spend hours doing so and still either miss things or mistake benign spots for malignancies. Typically, human experts can identify roughly 100 traits in a sample that might lead to a diagnosis. The AI, in this case, identified more than 1,000 individual traits that helped contribute to a diagnosis. It synthesized information, looked for patterns, and was extremely effective.
An AI bot isn’t going to replace a doctor. But it can give doctors the ability to diagnose issues faster and more effectively. This raises a further question. How important is subject-matter expertise to applying learning? Do you need deep background knowledge to be able to solve complex problems? Do you need to be a video game expert to create an expert video game playing AI? Or a trained pathologist to create an AI that detects cancer? This question will be tackled in the next blog post..