In this blog post, we’ll take a look at how AlphaGo and deep learning have been pushing the boundaries of artificial intelligence and how they work.
In March 2016, the AI Go program AlphaGo and world-class professional Go player Lee Sedol made headlines around the world. There have been AI vs. human matches before. For example, in May 1997, the chess AI computer Deep Blue defeated Garry Kasparov, the human world champion at the time. However, Go presents a different challenge to AI than chess. Chess is a game of moving pieces on an 8×8 board according to rules. Go, on the other hand, is a game of alternating stones on a 19×19 board, where the number of possible moves is 360 to the power of 10. That’s more than the number of atoms in the universe combined. For this reason, Go has long been considered unconquerable by artificial intelligence. In a 2009 interview, an authority on computer algorithms even claimed that no Go algorithm could beat a professional player within 100 years. But the results were shocking. AlphaGo won four out of five matches and sent shockwaves around the world. Machine learning, especially deep learning, played a key role in AlphaGo’s conquest of Go, a game that was thought to be impossible. But what is machine learning and deep learning, and how did it conquer previously uncharted territory?
Machine learning is a technology that enables artificial intelligence by allowing machines to learn from the data they are fed, rather than having to be programmed by humans. Machine learning can be broadly categorized into supervised learning, unsupervised learning, and reinforcement learning depending on the type of data being studied.
Supervised learning uses data where the desired outcome is specified for each piece of data. For example, in the case of image recognition AI, the computer teaches itself by providing data with desired outcomes, such as “this image is a cat,” “this image is a dog,” and so on. This way, when a new image comes in, the computer creates an algorithm that can tell if it’s a cat or a dog. The results are relatively accurate because it learns from pre-populated outcomes, but it requires a human to determine the outcome for each piece of data.
Unsupervised learning uses data that doesn’t have the desired outcome as input. Using the image recognition example again, a computer is taught to distinguish between dogs and cats by learning autonomously from data that doesn’t tell it whether the image is of a dog or a cat. This can be seen as a more advanced method than supervised learning, but it requires much more computation and is generally less accurate than supervised learning.
Reinforcement learning rewards the AI for the actions it takes in each state. The AI then trains itself in the direction that maximizes the reward. AlphaGo evolved through reinforcement learning. It used reinforcement learning by setting up a reward system that rewarded winning moves with (+1) points and losing moves with (-1) points. At each moment of the game, AlphaGo learned which actions had the highest probability of winning.
There are many different methodologies for learning from such data, including deep learning based on artificial neural networks. An artificial neural network is a learning algorithm that mimics the structure of the human brain, where artificial neurons form a network through synaptic connections and vary the strength of those connections to form an algorithm. In deep learning, the layers of this artificial neural network are deeply constructed, and the learning process proceeds through multiple neural networks.
The concept of deep learning has been around for a long time, but its practical use has been challenging. For example, it used to take three days to train an algorithm that could distinguish between 10 numbers. However, as computer power has increased dramatically, these speed issues have begun to be addressed. A major resurgence in deep learning came in 2012, when a deep learning algorithm won the ILSVRC, beating all the other algorithms in the competition. This sent shockwaves through the academic community, and deep learning has since become a mainstream part of machine learning and artificial intelligence.
Since deep learning became mainstream, the pace of progress in artificial intelligence has been unimaginable. At the 2015 ILSVRC, a Microsoft team demonstrated human-like image recognition with 96% accuracy. AlphaGo has continued to improve since its epic match against Lee Sedol and hasn’t lost a single match since. Nowadays, deep learning is even being used to create AI assistants that can talk to humans and act as secretaries. The time has come for deep learning to change the world.
Deep learning is now clearly at the center of global technology trends. NVIDIA, the company that produces the GPUs used in deep learning, is the sixth largest company in the world by market capitalization. While many people are amazed by AlphaGo and the performance of recent AI assistants, there’s more to come. The possibilities of deep learning are endless. If you’re an engineer, it’s worth diving into deep learning at least once. Even if you’re not an engineer, it’s fun to imagine what advancements will be made and what will be created with deep learning. With deep learning, your imagination can become reality.
