History of AI: The Evolution of AI Waves
It is widely believed that AI has experienced two waves and is now undergoing a third wave.
Below, we will review the history of AI and understand what AI really is.
What Exactly is AI—General AI vs. Narrow AI
No technology currently exists that can replicate human intelligence. DeepMind developed AlphaGo, and its founder, Demis Hassabis, aims to create a type of Artificial General Intelligence (AGI) that can make intelligent judgments on a variety of problems by merging machine learning with neuroscience.
While the technology now referred to as AI can perform at or above human levels in specific conditions like autonomous driving and board games, it lacks generality. We refer to this type of AI that is designed for specific tasks as Narrow AI.
Many of the technologies used in Narrow AI have been cultivated during the previous AI waves.
The definition of AI is somewhat vague because AI is not a single technology but a collection of multiple technologies.
People originally thought that General AI, Narrow AI, and other automation technologies should be discussed separately. However, technologies that were previously confined to automation or big data analysis are increasingly branded as AI in products and services due to the current AI wave.
The History of AI Waves
The term AI was coined during the first wave of AI. At the Dartmouth Conference in 1956, John McCarthy, an assistant professor in the mathematics department at Dartmouth College, named machines that could think like humans as “artificial intelligence.” Since one of the prototypes of modern electronic computers, ENIAC, was created in 1946, we can say that the concept of AI began to exist during the early days of electronic computing.
Researchers at that time attempted to achieve human-equivalent intelligence through techniques known as reasoning and exploration. However, they could only solve puzzles and some simple games, yielding almost no practical results. The subjects of research were idealized “toy problems” that had clear rules and constraints. Consequently, AI fell into a low tide by the 1970s.
The second wave of AI occurred in the 1980s. During this period, researchers taught machines expert-level knowledge as rules to develop “expert systems” capable of solving problems. Although there were successful commercial applications, the scope was limited. Eventually, the wave subsided again, as the difficulty of teaching rules to AI exceeded our expectations.
The practicality of cutting-edge machine learning is the driving force behind the current third wave of AI. Machine learning refers to the technology that enables computers to recognize sounds and images and make the most appropriate judgments by learning from large amounts of data.
This idea is not new; its prototype appeared during the first AI wave in the 1960s. However, it took a significant amount of time before machine learning reached a practical level because it requires a large amount of training data and substantial computing resources.
In the early 21st century, the cost of building big data was finally brought within acceptable limits, making it easier to acquire large amounts of training data.
The machine learning technology leading the third wave of AI actually encompasses multiple methods. The most notable among them is deep learning. Deep learning is a technique that uses a “neural network” that mimics the human brain to learn from large data sets.
The idea of “neural networks” actually dates back a long time, but the mainstream method for implementing deep learning emerged in 2006.
Figure 1 illustrates the development history of artificial intelligence (AI).
Figure 1: The Development History of Artificial Intelligence (AI)
Active Reasoning and Search Techniques
The AI research that began in the 1950s produced many methods and techniques, many of which are still in use today. For example, the search techniques that emerged during the first AI wave have undergone several evolutions and now, together with reinforcement learning, form the main technologies behind AlphaGo.
Before understanding what search is, let’s first look at a simpler game than Go: tic-tac-toe. In tic-tac-toe, we can use a search tree method to find out what the next move should be.
A search tree is a method that presents all possible options from a certain state in a tree format for searching. For instance, if the current state of the tic-tac-toe board is state A as shown in Figure 2, then there are 5 possible states for the next move.
In the leftmost state B, there are 4 possible states for the next move. In tic-tac-toe, if we count all the board states from the first move, there are a total of 5478 possible moves.
Therefore, if all moves are preemptively searched, we can easily handle all situations and thus win the game.
Figure 2: Search in Tic-Tac-Toe
However, the board states in Go and Japanese Shogi are extremely complex, making it difficult to win using this method.
In Go, there are 10 to the power of 170 possible board states. Such a vast amount of data is challenging even for supercomputers to handle. Therefore, for Go and Shogi, we generally can only adhere to a certain strategy and deeply explore only the moves that are likely to win.
AlphaGo employs Monte Carlo Tree Search technology. Monte Carlo Tree Search does not conduct a comprehensive search like previously described but randomly generates moves until the game is about to end. After repeating this process for each state, it focuses on searching the route with the highest probability of winning. By the way, the name Monte Carlo is derived from the Monte Carlo district in Monaco, known for its casinos. This name is often used when utilizing randomness and stochasticity in simulations and other methods.
Currently, active traditional techniques are not limited to search; there is also rule-based AI. This was the core technology of the second AI wave and is now widely applied in the thriving development of chatbots and voice terminals.
Additionally, one of the characteristics of the third AI wave is the increasing application of combining these traditional AI techniques and methods with deep learning. For example, reinforcement learning and generative model methods are not new technologies, but by combining them with deep learning, methods such as deep reinforcement learning and deep generative models have emerged.
The revival of traditional techniques after upgrades is also an important aspect of this AI wave that cannot be overlooked.