This blog post examines how AlphaGo and deep learning have advanced by breaking through the limits of artificial intelligence, and what the underlying principles are.
In March 2016, the match between the artificial intelligence Go program AlphaGo and world-class professional Go player Lee Sedol became a major global topic. Matches between artificial intelligence and humans had existed before. For example, in May 1997, the chess AI computer Deep Blue defeated Garry Kasparov, who was the human world champion at the time. However, Go presents a fundamentally different level of difficulty for AI implementation compared to chess. Chess is a game played on an 8×8 board where pieces are already placed and moved according to set rules. In contrast, Go is played on a 19×19 grid where players take turns placing stones anywhere, resulting in a total number of possible moves reaching 10 to the 360th power. This number is vastly larger than the total number of atoms in the universe. Consequently, Go was long considered an insurmountable frontier for artificial intelligence. In a 2009 interview, an authority in computer algorithmics even asserted that an algorithm capable of defeating a professional Go player would not emerge within 100 years. Yet the outcome was shocking. AlphaGo won four out of five matches, causing a global sensation. The key to AlphaGo’s conquest of Go, once deemed nearly impossible, was machine learning, particularly deep learning. So what exactly are machine learning and deep learning that enabled the conquest of a domain previously thought unattainable?
Machine learning refers to the technology where artificial intelligence is implemented not by humans meticulously programming each algorithm, but by machines learning autonomously through input data. Machine learning can be broadly categorized into supervised learning, unsupervised learning, and reinforcement learning based on the form of the data being learned.
Supervised learning uses data where the desired outcome is explicitly specified for each data point. For example, in image recognition AI, data containing the desired outcome is provided, such as “This image is a cat,” “This image is a dog,” enabling the computer to learn independently. Through this, the computer creates its own algorithm capable of distinguishing whether a new image is a cat or a dog. Because it learns from pre-input results, the outcomes are relatively accurate, but it requires the cumbersome task of manually assigning results to each data point.
Unsupervised learning uses data without predefined desired outcomes. Returning to the image recognition example, it involves the computer learning autonomously to distinguish dogs from cats using data lacking information about whether an image depicts a dog or a cat. While this can be seen as a more advanced method than supervised learning, it requires significantly more computation and generally yields lower accuracy compared to supervised learning.
Reinforcement learning provides rewards for actions taken by the AI in each state. The AI then learns autonomously to maximize these rewards. AlphaGo developed through this reinforcement learning. Rewards were set such that winning a game earned (+1) points, while losing resulted in a loss of (-1) points. AlphaGo learned the action with the highest probability of winning at every moment during the game.
Various methodologies exist 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. Artificial neurons form a network through synaptic connections, and the algorithm is constructed by adjusting the strength of these connections. In deep learning, these artificial neural networks are configured with deep layers, enabling learning to progress through multiple neural networks.
The concept of deep learning emerged long ago, but practical implementation was challenging. For instance, training an algorithm to distinguish just 10 digits could take up to three days. However, as computer performance significantly improved, these speed issues began to be resolved. The major catalyst for deep learning’s resurgence was the 2012 ILSVRC competition, where a deep learning algorithm overwhelmingly outperformed all other existing algorithms to win. This caused a major shock in academia, and since then, deep learning has become the dominant trend in machine learning and artificial intelligence.
Since deep learning became mainstream, the pace of AI development has been beyond imagination. At the 2015 ILSVRC, a Microsoft team achieved 96% accuracy, demonstrating image classification 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 roles have been created through deep learning. The era of deep learning changing the world has arrived.
Deep learning has now clearly become the core of global technology trends. NVIDIA, the company producing GPUs used for deep learning, has risen to become the world’s sixth-largest company by market capitalization. While many 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. Engineering students should definitely dive into deep learning at least once. Even non-engineers can enjoy imagining what advancements deep learning will bring and what it will create. With deep learning, your imagination might just become reality.
