3 Types of Machine Learning
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
In supervised learning, we train an agent using labeled data. That’s input data we know the answer for, but don’t show to the machine until after it’s made a guess. We can train a machine to process and output a decision that matches the expected output as accurately as possible. By feeding a machine inputs over and over and having a penalty for getting the wrong answer, we can nudge it toward the correct answer. Once the machine is properly trained, it can give the right answer every time to match our expected output.
Consider this real-life example: You head to the library looking for a book about AI and meet a librarian who is a huge tech geek. He knows the best books on AI. To get you to learn more about supervised learning in real life, he makes you look for the books by yourself and only tells you if the book is about AI or not. Through this trial and error process, you are eventually able to find good books on AI and can now probably know enough to help other people pick good books about AI as well. The librarian is like the labels on the data, they always have the right answers, but you only get feedback after making a guess.
Unsupervised learning is super powerful in the real world. It’s what enables services like Google photos to automatically categorize the people in your pictures into albums. Or for Facebook and Google to send you targeted ads by analyzing what people with similar interests or web activity as you buy. They don’t actually listen to your conversations or tap into your dreams; it’s just AI (which might actually be scarier.)
In unsupervised learning, we have to get machines to generate useful output for problems we don’t already have the answer to. Machines try finding trends and patterns when given unlabeled data as input. The process of recognizing patterns and trends is how a learning algorithm is able to make sense of the data.
Put yourself in this scenario: You bought a huge box of mystery chocolate from a post-Halloween sale. It’s all unwrapped and tossed into a giant box so you decide to make chocolate bouquets for some friends. But you have no idea what type of chocolate you’re getting so you randomly try a whole bunch. Eventually, you’re able to recognize a piece of chocolate just by looking at it and can easily tell what brand it is and if it has nuts or caramel inside. You could then easily group them into categories to give to your picky friends who only eat certain types of chocolate.
You might’ve heard about Elon Musk creating an AI bot to beat the world’s best Dota 2 players or Google’s DeepMind AlphaGo Zero project. Both were built using reinforcement learning algorithms. It’s insane to think that artificial systems can not only learn how to play our games better than we can but actually teach us better ways of playing them ourselves. Remember before when I said machine learning was just about receiving an input, processing it, and returning useful output? If that’s all there is to it, then where does reinforcement learning fit in?
Reinforcement learning follows a basic idea of teaching a system how to minimize punishment and maximize reward. It’s somewhat like supervised learning with the idea of penalizing the machine for getting the wrong answer. Essentially, the agent learns through trial and error in order to get an answer. The difference is that we don’t know what the right answer looks like. There’s no set input/output combination for reinforcement learning. The algorithm only knows how to achieve a goal based on how you train it and what you set as rewards and punishments.
Imagine you’re running through a maze blindfolded. The walls are made of spikes and there’s a hospital on the other side. You have to make your way through only knowing which way to go based on if you run into a wall or not. The first few times you’ll bump into lots of walls and incur a plethora of injuries. It will hurt. A lot. But each time you complete the maze, you get treated for your injuries. After lots of repetition, you’ll be a lot better at navigating through the maze while avoiding the walls. Then you can just sit at the hospital and eat all their banana pudding. Because you deserve it. You successfully learned how to get through the maze unscathed with no outside help. Just that walls are bad and hospital pudding is good.
Industries Affected By Machine Learning
All of them. Next!
Seriously. The amount of data in the world is growing at an exponential pace and right now, it seems like machine learning is the most efficient way to parse all of it (besides potentially using quantum computers.) It’s what machine learning was built for: Input -> Analysis -> Output. There’s literally an entire field dedicated to the processing and extracting knowledge from data (conveniently called data science.)
It seems plausible that in the near future AI will be like electricity. Every business in every industry will be using it within the next 10 years or they’ll fall behind real fast. It’s like how today you don’t ask corporations if they use electricity or not; they just have to.
- Machine learning at its core is the taking an input, analyzing it and producing a useful conclusion or decision
- Supervised learning uses labeled data and trains an algorithm to match an expected output given a certain input
- Unsupervised learning works with problems we don’t already know the answer to by clustering and sorting inputs based on certain patterns and trends
- Reinforcement learning doesn’t really aim for a traditional output per se since it works off of solely an incentive system where the goal is to minimize negative penalties and maximize positive reinforcement