In this blog post, we’ll explore how AlphaGo and deep learning have broken through the limits of artificial intelligence and evolved, as well as the principles behind them.
In March 2016, the match between the AI Go program AlphaGo and world-class professional Go player Lee Sedol became a major global sensation. There had been matches between AI and humans before. One example is the event in May 1997 when the chess AI computer Deep Blue defeated Garry Kasparov, who was the reigning world champion at the time. However, Go presents a level of difficulty in AI implementation that is on a completely different scale from chess. Chess is a game played on an 8×8 board where pieces are moved according to set rules. In contrast, Go is a game played on a 19×19 board where players take turns placing stones anywhere on the board, resulting in a total of 10 to the 360th power possible moves. This figure is far greater than the total number of atoms in the entire universe. For this reason, Go has long been considered a realm beyond the reach of artificial intelligence. In a 2009 interview, a leading authority in the field of computer algorithms even asserted that a Go algorithm capable of defeating a professional player would not emerge within 100 years. Yet the outcome was shocking. AlphaGo caused a sensation worldwide by winning four out of five matches. Machine learning, particularly deep learning, played a pivotal role in AlphaGo’s conquest of Go—a feat once considered nearly impossible. So, what exactly are machine learning and deep learning, and how did they enable the conquest of a realm previously deemed out of reach?
Machine learning refers to a technology in which artificial intelligence is implemented not by humans manually programming every algorithm, but by machines learning on their own through inputted data. Machine learning can be broadly categorized into supervised learning, unsupervised learning, and reinforcement learning, depending on the form of the training data.
Supervised learning uses data where the desired outcome is explicitly specified for each data point. For example, in the case of image recognition AI, data containing desired outcomes—such as “This image is a cat” or “This image is a dog”—is provided so that the computer can learn on its own. Through this process, the computer creates its own algorithm capable of distinguishing between cats and dogs even when presented with new images. Since it learns from pre-inputted results, the outcomes are relatively accurate, but there is the inconvenience of having to manually define the results for each data point.
Unsupervised learning uses data that does not have the desired results pre-inputted. Returning to the image recognition example, this involves the computer learning autonomously from data that lacks information on whether an image is a dog or a cat, enabling it to distinguish between the two. While this can be viewed as a more advanced method than supervised learning, it requires significantly more computational power and generally yields lower accuracy compared to supervised learning.
Reinforcement learning provides rewards for the actions taken by the AI in each state. The AI then learns on its own to maximize these rewards. AlphaGo developed through this reinforcement learning. It conducted reinforcement learning by setting rewards such that winning a game earned a score of (+1) and losing resulted in a loss of (-1) points. AlphaGo learned the action with the highest probability of winning at every moment of the game.
There are various methodologies for learning from such data, and deep learning based on artificial neural networks is one of them. 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 adjust the strength of those connections to construct the algorithm. In deep learning, these artificial neural networks are structured in deep layers, allowing learning to proceed as data passes through multiple layers.
Although the concept of deep learning emerged long ago, it was difficult to apply in practice. For example, it used to take as long as three days just to train an algorithm to distinguish between ten digits. However, as computer performance improved significantly, these speed issues began to be resolved. The catalyst for the massive resurgence of deep learning was the 2012 ILSVRC competition, where a deep learning algorithm overwhelmingly outperformed existing algorithms to win the championship. This sent shockwaves through the academic community, and since then, deep learning has established itself as the dominant trend in machine learning and artificial intelligence.
Since deep learning became the mainstream, the pace of AI development has been beyond imagination. At the 2015 ILSVRC, the Microsoft team achieved 96% accuracy, demonstrating image recognition capabilities on par with humans. AlphaGo has continued to evolve since its match against Lee Sedol and has never lost a single game to date. Recently, AI assistants capable of conversing with humans and performing secretarial tasks have even been developed through deep learning. We have entered an era where deep learning is transforming the world.
Deep learning has now clearly become the core of global technology trends. NVIDIA, a company that manufactures GPUs used for deep learning, has risen to become the sixth-largest company by market capitalization worldwide. While many people were amazed by AlphaGo and are impressed by the performance of recently released AI assistants, even more astonishing developments are yet to come. The possibilities of deep learning are limitless. If you’re an engineering student, it’s worth diving into deep learning at least once. Even if you’re not an engineering student, it’s fun to imagine what advancements will be made and what will be created thanks to deep learning. With deep learning, your imagination might just become reality.
