What is it that Machine Learning Can Actually Do?

Machine Learning (ML) is getting a tremendous amount of attention lately, as it is the fastest-growing area of artificial intelligence. Last time I explained why it’s a big deal – because instead of giving a computer instructions, we give it examples, and the computer learns in the same way that people learn: through experience. This enables computers to do things that they have never been able to do before. But to understand what kind of things, it’s helpful to understand the basic categories of ML.

Kinds of Machine Learning

You may have heard of some of the algorithms used by ML to do its thing – Support Vector Machines, Deep Neural Networks, Naïve Bayes, K-means, etc. But unless you’re aiming to be a data scientist, you probably don’t need to know how these algorithms work. What is helpful to know is that the algorithms used in ML fall into one of three categories of learning: Supervised, Unsupervised, and Reinforcement.

1.    Supervised learning: the computer learns by example. For example, give the computer thousands of pictures and for each one tell it “this is a cat,” “this is a dog” and eventually it is able to tell if a new picture is of a cat or of a dog.

2.    Unsupervised learning: from a lot of content, the computer attempts to discover something meaningful on its own. For example, give it thousands of pictures of cats and dogs, and tell it to separate all the pictures into two groups. Chances are good it will separate them into cats and dogs, and learn distinguishing characteristics of each so that it could tell if a new picture is of a cat or a dog.

3.    Reinforcement learning: in a dynamic environment (like a game), the computer is given a goal and is scored based on how well it achieves that goal. After a lot of trial-and-error, it starts to learn techniques to better achieve the goal. For example, give a computer the dimensions of 17 different items to pack into a box for shipping, and the smaller the dimensions of the final arrangement, the better the score; it will learn how to best pack the items to minimize space.

Unsupervised Learning

To date, unsupervised learning has been the least used, but . Since computers don’t have “common sense” or intuition, the results of unsupervised learning are somewhat unpredictable.  The result might be an effective classifier or a new discovery, or it could be something useless. It’s kind of like giving your 5-year old free reign in the kitchen with the instructions to “cook something for dinner.” You’ll get a result – and he’ll be able to tell you what he had in mind – but it might not be edible.

Reinforcement Learning

Reinforcement learning is valuable in dynamic environments where multiple actions must be taken over time. It’s especially effective when the quality of that action can be measured by the computer itself (e.g., minimum distance between two points, win the most games of tic-tac-toe, get the highest score in Pac Man). It’s very well-suited for games and robotics. It’s used by UPS to set delivery routes for today’s packages, Google to beat the world champion in the incredibly complex game of Go, and Tesla to fine-tune its autopilot.

Supervised Learning

Of the three, supervised learning is the most widely used in business today.  It’s used by Google to improve Internet search results (and Google Translate), e-mail providers to filter out spam, examine pathology to diagnose cancer (and soon radiology), Netflix to recommend movies, Amazon to suggest items to purchase, Uber to set fares and estimate times, PayPal to identify fraudulent transactions, wireless carriers to anticipate who might switch to another provider, IBM to compose songs, insurance companies to guide underwriting, to improve customer service and satisfaction, and even the penal system to recommend prison sentences.

Not Mutually Exclusive

These learning techniques are frequently used together; in fact, some of the best results come with hybrid approaches. That’s what enabled such rapid progress with AlphaGo, Google’s world-champion Go player. It started with supervised learning, training the computer on “good” and “bad” moves by showing it actual games from tournaments played by experts. Then, once the computer had a solid understanding of how to play the game, it used reinforcement learning to play itself over and over again, adding to its experience faster and improving its strategies – much faster than it ever could have by playing humans.

Training Takes Time

Just as with human learning, ML takes time to get good at what it does. The more complex the problem or activity, the more training is required.  That’s one of the reasons ML hasn’t taken off even though most the algorithms have been around for decades (and some nearly a century). Every year computers get faster and are capable of handling more data, and the recent advent of parallel computing frameworks (and availability of data) has really allowed ML to see much broader use.

Even with today’s capable platforms, ML training can still take a long time. Just as it takes a musician years to master his instrument, it’s not unusual for ML training to take months.  ML can also take longer because it can be an iterative process where results from initial training efforts are used to fine-tune parameters, and then training starts over with new settings.

Future

Now that we have a better understanding of ML and where it works and doesn’t, we can make predictions about what it will do in the future. Next time, I’ll share some thoughts about where it’s headed.