This blog post examines how close artificial intelligence can come to human intuition through the showdown between AlphaGo and Lee Sedol.

Lee Sedol vs AlphaGo: The Clash of Human and AI

One of the biggest stories of 2016 was the match between Lee Sedol 9-dan and Google DeepMind’s AlphaGo. Lee Sedol was hailed as the strongest player in the Go world, while Google DeepMind was a subsidiary of Google, one of the world’s most innovative IT companies. The clash between these two entities transcended a simple Go match, forming a traditional yet fascinating ‘computer vs. human’ rivalry that sparked significant public interest. This match was seen as an opportunity to test the existing limits of artificial intelligence, sparking curiosity about the boundaries of both the technology and human capability.
The match consisted of five games, and ultimately, AlphaGo defeated Lee Sedol 9-dan with an overwhelming score of 4:1. After the matches concluded, people came to believe that defeating a computer capable of rapidly calculating all possible moves was virtually impossible for humans. Although Lee Sedol won only one game, many hailed him as a monumental challenger and paid tribute to his achievement. If AlphaGo had relied solely on calculating all possible moves, this victory would have been merely the result of a computer with improved computational power. However, AlphaGo’s victory stemmed not from hardware performance improvements, but from revolutionary advancements in its internal algorithms. This demonstrated that AlphaGo’s understanding of Go was fundamentally different from that of previous AIs. AlphaGo’s algorithms have since been applied across multiple fields, opening the possibility of profoundly impacting our lives.

Existing Algorithms: The Minimax Algorithm and Its Limitations

To understand why AlphaGo was special, we first need to understand the Minimax algorithm used by previous board game AIs. In games like Connect Four or chess, AI had already been capable of defeating humans for a long time; in chess, it had been decades since a computer first defeated a world champion. The Minimax algorithm used at the time was based on the concept of ‘considering all possible moves and selecting the best one’. True to its name, the Minimax algorithm adopts a strategy of preparing for the worst possible outcome the opponent might make while deriving the best possible result for itself. On a chessboard, with its limited 8×8 grid and specific piece movements, the number of possible moves is relatively small. With sufficient computational power, it can look as far ahead as possible and make moves accordingly.
However, Go presents a completely different scale of possibilities. The Go board spans a vast 19×19 grid, and with almost no restrictions on where stones can be placed, the number of possible moves reaches hundreds of millions or even trillions from the very first move. If one attempted to calculate all possibilities using the Minimax algorithm, considering just six moves ahead would require analyzing approximately 22 trillion cases. Assuming one calculation per second, completing all computations would take 700,000 years. Even chess proved too complex to analyze all possibilities through simple computation alone, necessitating various shortcut techniques. However, these approaches were no longer viable for complex games like Go.

AlphaGo’s Innovation: Intuitive Judgment Through Deep Learning

AlphaGo’s victory over humans in Go stemmed from introducing a novel approach that overcame the limitations of the Minimax algorithm. Because selecting the best move is extremely difficult in Go, AlphaGo employed a method that drastically reduced computational processes by utilizing judgment similar to human intuition. The core of this new approach came from a technique called Deep Learning. Deep Learning is an artificial intelligence methodology that mimics the neural networks of the human brain, enabling it to solve complex problems through learning and intuition.
Deep learning achieves learning by processing input data through multiple layers of neurons in artificial neural networks. In this process, AlphaGo predicted moves with high win rates at specific positions based on learned patterns, rather than simply calculating every possible move, and then executed those moves accordingly. Through gradient descent, one of deep learning’s key learning techniques, the AI reduces errors through repeated learning, enabling more accurate judgments. This deep learning technique has the advantage of performing all operations on a matrix basis, allowing for parallel processing. This enables calculations to be processed at tremendous speeds using GPUs.

The Introduction of CNN: Recognizing the Go Board as an Image

Among the deep learning techniques AlphaGo employed, CNN (Convolutional Neural Network) played a pivotal role. Originally an algorithm demonstrating great success in image recognition and classification tasks, CNN excels at recognizing patterns and extracting features within images. CNNs operate by analyzing each pixel in an image to classify the shapes and features of objects contained within. AlphaGo replaced the Go board with CNN pixel data, visualizing the placement and configuration of Go stones for learning. Based on CNN features, AlphaGo could analyze each move on the Go board like analyzing color patterns in an image, calculating the probability of winning for the next move. This enabled AlphaGo to make highly intuitive moves through learning and pattern recognition, not just simple calculations, allowing it to make decisions on a different level than existing AI.
AlphaGo incorporated a self-learning function distinct from existing AI, developing judgment capabilities closer to human intuition by repeatedly training itself using game records. This signifies that deep learning technology, by granting AI self-learning capabilities, enables it to perform tasks requiring complex thought processes, going beyond merely solving given problems. This self-learning capability demonstrates that AI is approaching a stage where it can increasingly mimic human thought processes and independently generate new strategies.

The Future Unlocked by Deep Learning: Artificial Intelligence Entering Our Lives

The impact of deep learning techniques on daily life is beyond imagination. Even now, artificial intelligence provides significant assistance by mimicking human thought and judgment in diverse fields such as image classification, language translation, and speech recognition. As deep learning becomes more widespread and AI learns to recognize an individual’s behavioral patterns, it will evolve beyond a simple assistant. Instead, it will predict a person’s actions to reduce errors and autonomously make necessary decisions. For instance, if an era arrives where AI learns a person’s health status and daily habits to detect risk factors in advance, or predicts traffic conditions in real-time to guide the optimal route, our lives will become more convenient and safer.
Furthermore, as AI becomes deeply involved in actual human life and gradually takes on larger roles, we will experience new forms of living unimaginable in the past. In this sense, the 2016 AlphaGo vs. Lee Sedol match transcended a simple Go game; it was the starting point where the boundary between humans and AI began to dissolve, serving as a preview of our future coexisting with AI.