The first definition of machine learning dates back to 1959, from American AI pioneer Arthur Samuel:
Machine learning is the field of study that gives computers the ability to learn without being explicitly programmed.
The key elements here are learning and without being explicitly programmed. Let’s focus on the latter first. Explicitly programming a computer means defining the rules and instructions it must follow to perform a specific task. This is what software engineers do when they write software that handles your everyday tasks like doing taxes or filling out spreadsheets.
People without programming experience often feel like software engineers are powerful creatures who can bend machines to their will. Unfortunately, things are not always that easy. Try to think about the various decisions you make as you perform some trivial actions: can you explain the process you follow to recognize your friends when you see them? All the split-second decisions you make while driving? Can you list all the English grammar rules you apply as you talk? If you can’t precisely explain how you do something, there’s no chance that you can instruct a computer to do it.
Samuel proposed to replace “instructing computers” with “giving them the ability to learn.” If you think about it, learning instead of following instructions is (coincidentally?) what human beings do all the time. Our mothers and fathers don’t teach us their native tongue by giving us grammar books at the tender age of one. They just speak to us naturally, and we learn from their example, applying thousands of grammar rules without even knowing it. In fact, our brain is capable of automatically extracting rules way before it becomes capable of rationally understanding grammar
at school! Even for us humans, it looks like learning rules from examples can be easier than being told about them.
In the same way we learn from experience, machine learning (ML) techniques allow computers to learn from data. Let’s make it more concrete with a classic toy example: teaching a computer to tell dogs from cats in pictures. If you had to teach a kid to perform this task, you wouldn’t pick up a veterinary book and start reading about the differences in ear shape or fur color. Instead, you’d probably just point them to a few pictures and let their brain do its magic.
An ML solution to the “dog or cat” problem is similar to our childhood learning experiences. We feed the computer thousands of images of cats and tell it “these are cats,” and then thousands of images of dogs and tell it “these are dogs.” Finally, we let it figure out the difference between the two pets automatically. We don’t have to explain the key elements that distinguish dogs from cats. A good ML application learns to figure that out from the examples it receives.
You may start to sense why ML couldn’t possibly have blossomed before the 2000s. The main ingredient of this set of techniques is data, and the internet has made collecting data much easier. The other crucial ingredient for ML is computing power: learning from data doesn’t happen for free, and computers need fast processors to perform this task. Thanks to cloud computing and increases in processing power, access to powerful computers has never been so easy and cheap.
To give you a sense of how much things have changed in just a few years, we asked Alex Waibel, one of the pioneers of AI in speech recognition and among the first hires of Facebook’s AI team, how different it was to work on ML 20 years ago. The most powerful computer he could use in the early 2000s was as big as an apartment, cost a few million, and he needed to rent it to train his models. Today, he has much more computing power sitting on his desk for a few thousand dollars. Your phone is probably more powerful than what top researchers had available just 20 years ago.
Availability of data and cheap computing power created the perfect environment for machine learning to bloom. Indeed, many (most) of the coolest consumer-facing applications of what we call AI today rely heavily on ML: the Siri voice virtual assistant, Google Translate, self-driving cars, and many more.
Going back to the history of AI, it seems that ML is the engine that powered today’s AI explosion, finally bringing some hope after the last AI winter of the 1980s.
In fact, the success of modern AI has been so dependent on ML techniques that people are often confused about the difference between the two. What is artificial intelligence, then? Let’s find out in this blog post.