Machine Learning For The Rest Of Us
A brief, dead simple explanation of Machine Learning — what is it, where does it come from and why is everybody talking about it?
Introduction
Everybody seems to be talking about this right now: Machine Learning. Actually, it’s usually a mix of different buzzwords, mixing in things like Deep Learning and Artificial Intelligence. But what do these things mean? What’s the reason why people seem to go crazy over this right now?
These are the questions I try to answer in this article. I will explain what Machine Learning is and why there’s so much buzz around it. To begin with, we first have to look at the big picture: AI.
Artificial Intelligence
Let’s have a quick “reality check” here. What do you think most people have in mind when they hear the term “AI”? When I talked about this with colleagues and friends, most of them thought about something like this:
That’s not AI. Thanks to Hollywood, people tend to think about mad robots trying to kill us when they hear the term “AI” (or they think about friendly robots, usually with a funny character flaw). But it’s important to understand that an artificial intelligence does not have to be a robot. In fact, it doesn’t need to have a physical body at all. This is difficult for us, because we tend to anthropomorphize things, especially if they appear to be intelligent.
Luckily, we already have lots of examples of existing AI. Some of them are with us most of the time, right in our pocket. Our smartphones contain various AIs which help us to navigate our surroundings, recommend products we did not know we wanted or kick our ass in chess.
See? No robots at all. AI is the brain, not the body. AI ≠ robots.
We also see that we tend to not call things AI anymore once we got used to them. As long as it seems to be smarter than anything else in the field and very experimental, it’s AI. As soon as you and your mum can buy it and it becomes part of your life, it’s just another tech product.
To put things into perspective, experts have defined three levels of artificial intelligence:
- Artificial Narrow Intelligence (or ANI): An AI that can do one thing well (better than a human), but is horrible at pretty much everything else. This is what I just talked about. The maps app on your smartphone, the chess computer or the e-commerce system that finds products related to your purchase are all examples of ANI.
- Artificial General Intelligence (or AGI): An AI that is as smart as a human, not only in regards to one specific task, but across the board. This is not here yet, but it’s the reason why all the big players like Google, Apple and Facebook are putting the focus of their research and development to AI. They want to get there soon and they want to be the first one doing it.
- Artificial Superintelligence (or ASI): At this stage, an AI is much, much smarter than any human being and can do all sorts of crazy things. It’s very exciting and scary to think about this, but that’s out of scope for this post.
As of today, every existing AI is a narrow one, but each improvement brings us one step closer to AGI and, ultimately, to ASI. As we will see shortly, this is one of the “Whys” behind the whole Machine Learning thing — getting closer to AGI.
The Road To AGI
Did you ever wonder how you actually teach something to a machine? I mean, not the actual code, but rather the principles and ideas behind that? Let’s assume we want to teach a machine how to recognize a cat. For us humans, this is pretty much as simple as it gets — everyone would instantly recognize a cat, right?
It turns out that this is a pretty difficult task for a machine.
Until recently, the approach to teach something like recognizing cats in images to a machine meant you would list all the things that make up a cat, like whiskers, fur, two eyes and four legs. The tricky part about this is that we humans do not think about these things consciously — we look at an image and we immediately decide whether or not it shows a cat. It’s a visceral, subconscious decision and that’s why it’s so difficult to articulate.
It becomes even more complicated and tedious for things that are difficult for us, like playing chess. In such a case, you first need to get a good understanding of how it works yourself before you can teach it to a computer. That takes a lot of time, especially considering the fact that you not only want a machine to be good at chess (or recognizing cats, for that matter), but one who is superior to every human on the planet.
To sum it up: The “old” approach of teaching something to a machine was trying to figure out how experts behave or, more broadly, how humans behave and quantifying that as rules to be told to the machine. Considering the fact that, to reach AGI, the machine would have to learn a lot of things that are very difficult to explain (like emotional intelligence or creativity), this approach doesn’t look very promising in regards to AGI.
Now think about how you learned to recognize a cat. Did your parents explain to you all the different elements that make up a cat? Most likely not. Instead, when you saw a cat for the first time, you were told that this is what people call a “cat”. The next time you saw a (different) cat, you may have already recognized it because it looked like that other thing you were told is called a cat. This is a prime example for one of our biggest strengths: pattern recognition. Your brain learned what makes up a cat, but not because you were told so. It did so by looking at sample data, making a guess, getting feedback and learning from that.
Recognizing Cats At Scale
That is what the Google Brain Project did back in 2012. They fed 12 million YouTube video thumbnails to a computer system and let it browse the data. They did not tell it to look for cats or what makes up a cat. In fact, they didn’t tell it anything about cats at all. At the end of the experiment, the system was capable of recognizing cats at a very good accuracy. It basically invented the concept of a cat on its own. Even more, not only did it learn to identify cats in an image, but also human faces and human bodies. The system learned to categorize the inputs without any guidance from an expert and without any rules.
How was that possible? Deep Learning. The basic idea behind Deep Learning is to model the data and data structure after the human brain and then feed it into a system equipped with algorithms enabling it to learn from that data.
Why “Deep”, though? Well, when modelling the data and its structure after the human brain, you’re in fact creating a neural network. The way the brain learns is that it strengthens the connection between neurons that lead to “good” results and weakens those connections that lead to “bad” results. A neural network mimics this process. It consists of many layers (that’s were “deep” is coming from) and each layer consists of many nodes (the “neurons”). The network is then going to train itself, given the input data, to set the connection strength between each of the nodes.
This process is the heart of the revolution: Instead of interviewing human experts, extracting their knowledge and procedure and then trying to teach a computer to mimic that, we feed a bunch of data into the computer and it will learn to classify by itself.
That is Deep Learning, a set of techniques where machines learn from data. But, there are other methods allowing machines to learn and to “become more intelligent”. They are all summarized under the term “machine learning”.
Deep Learning = a set of techniques within machine learning, where machines learn from data.
Machine Learning = a set of techniques that gives computers the ability to learn without being explicitly programmed.
Deep Learning > Machine Learning > AI
Now it’s time to put everything together and get back to the big picture. Most examples of current top-notch AI (AlphaGo, Watson, self-driving cars, etc.) use Machine Learning. It is also used in more “everyday” products like Google Photos automatically grouping pictures.
The reason why Machine Learning and Deep Learning are so hot right now is because it’s the first time in history we have sophisticated computing power on the one side and huge amounts of data to learn with on the other. Both of these developments allowed for Machine Learning to become a thing and as of now, it’s the most promising approach to artificial general intelligence.
Further Reading
Whew, we’re done! First of all, thanks a lot for reading — if you liked it, I’d appreciate a ❤.
Second, here’s a list of resources that helped me get a better understanding of the topic as well as further implications and possible future developments.
- Presentation from Andreessen Horowitz that gives a great summary of the whole topic, with historical background and examples all in a 45 minute video:
- Short and well-written 6 minute introduction to AI by Snips with good explanations of the most common terms:
- Ever wondered how AI might influence your job in the future? Here’s a great story from Rand Hindi that also touches the general topics of AI and Machine Learning:
- Another very good article that explains the general topics. This one is from Sam DeBrule and he’s not only providing a lot of other resources at the end of his post, but he is also running a newsletter on the topic (http://subscribe.machinelearnings.co/).
