2. The Leap of Machine Learning and Deep Learning
2024 will be recorded as an important turning point in AI history. Both the Nobel Prize in Physics and Nobel Prize in Chemistry were awarded to AI research, showing that AI has positioned itself as a tool that changes the fundamental paradigm of science beyond simple engineering innovation.
These achievements can be said to be the fruit of approximately 70 years of research and development since the concept of artificial intelligence first appeared at the Dartmouth Workshop in 1956. Early AI researchers sought to create machines that mimic human ways of thinking, but today AI is advancing to stages that transcend human intellectual capabilities. AI is now transforming not merely into a simple automation system but an intelligent entity leading new scientific discoveries, and the meaning of these changes is deepening.
This chapter will comprehensively examine AI's evolutionary process, from early neural network models to transformer-based latest technologies, and AI's industrial applications and ethical challenges.
1. The Concept of Artificial Intelligence and Early Development
1) The Definition of Artificial Intelligence and the Significance of the Dartmouth Workshop
Artificial intelligence (AI) is defined as technology that mimics human intelligent behavior and through this solves problems or makes decisions. However, this definition is very comprehensive and abstract, making it difficult to fully explain the essence of AI. Because artificial intelligence is not simply a single technology but a collection of wide-ranging technologies that make computers and machines operate intelligently.
The term AI was first officially used at a summer workshop held at Dartmouth College in the United States in 1956. This workshop was a gathering of 10 pioneering scholars including John McCarthy, Marvin Minsky, Claude Shannon, and Herbert Simon, who discussed methods for developing machines capable of mimicking human learning and problem-solving abilities. This workshop is evaluated as the starting point of modern artificial intelligence research and subsequently had a significant influence on the direction of AI development.
The core goal discussed at the Dartmouth Workshop was to explore methods for implementing human intelligent activities in machines. Scholars sought to research whether machines could deal with complex problems that humans solve through learning and reasoning. In particular, focus was placed on developing algorithms for problem solving, data-based learning methods, and giving machines logical reasoning ability.
The importance of the workshop lies in having systematically explored AI's possibilities for the first time. Through this, it presented the vision that computers could perform intelligent tasks unique to humans beyond simple calculation. As a result, AI research began developing based on diverse academic backgrounds including mathematics, psychology, and linguistics, and AI research established itself as a challenge to transcend the limits of technology.
Early AI research produced a number of noteworthy achievements. For example, programs that proved mathematical theorems, robot simulations performing commands in a geometric block world, and early language translation systems were developed. However, the technology of the time faced limitations in scalability for actual application fields due to limitations in computer performance and algorithms. This became an opportunity for AI research to recognize the need for more powerful computational power and new methodologies.
The Dartmouth Workshop became an essential turning point in materializing AI's future beyond being a simple venue for academic discussion. This gathering announced to the world that artificial intelligence could become a key to understanding human thinking alongside technological challenges. These efforts became the foundation leading AI's evolution, and are still evaluated as the foundation of artificial intelligence research today.
2) Early Cases and Technical Limitations
Early artificial intelligence research began with attempts to mimic human thinking and learning processes, exploring its possibilities through various application cases. During the 1950s–70s, AI researchers focused on implementing problem-solving and learning capabilities using computers, achieving several meaningful results. However, the technical achievements of this period simultaneously revealed significant limitations, providing important turning points in the evolution of artificial intelligence.
Early Cases
Early AI research began with developing programs specialized for specific problem solving. For example, programs that proved mathematical theorems succeeded in proving existing theorems more simply or deriving new theorems. Also, in virtual environments using geometric block worlds, robot arms could perform block stacking tasks according to commands input by users in natural language. These studies showed that computers could perform tasks beyond simple calculation.
In the 1970s, AI research expanded to various fields. For example, music composition programs that mimicked classical composers' styles partially implemented human artists' creative processes, and autonomous vehicle programs were designed to operate vehicles on their own based on environmental data. In addition, programs solving patent inventions or visual analogy problems were developed, experimentally proving AI's possibilities.
Technical Limitations
Early cases showed interesting possibilities, but revealed serious limitations in practical scalability and practicality. The biggest problem was the phenomenon of 'combinatorial explosion.' This caused systems to operate inefficiently or stop due to limitations of algorithms that had to explore all possible cases as problems became more complex. Problems of systems failing to operate properly in complex situations occurred frequently.
Also, early AI systems lacked the ability to handle uncertainty. Algorithms of the time presupposed all data being provided in clear and complete form, but in the real world incomplete data was much more common. Accordingly, in uncertain or changing environments, performance of AI systems deteriorated sharply.
Hardware performance limitations were also an important problem. Computers of the 1970s had extremely limited memory and calculation speed compared to the present, creating constraints in running complex algorithms. These hardware limitations limited AI research's scalability, making practical applications difficult.
Need for New Approaches
To overcome these technical limitations, researchers began exploring methods utilizing 'Heuristic Search' or 'Flexible Abstract Representation.' This was an attempt to develop more efficient search and problem-solving methods based on domain knowledge and experience rather than simply exploring all possible cases. However, these approaches also hit limitations in environments where sufficient data was initially difficult to secure.
As such, early AI research proved through interesting cases that computers could perform roles beyond simple calculation. However, at the same time, problems like combinatorial explosion, lack of uncertainty handling ability, and hardware performance limitations obstructed AI's practical development. These limitations became opportunities to explore new algorithms and technical approaches, becoming the foundation for subsequent AI development leading to neural networks and machine learning.
2. The Leap of Machine Learning and Deep Learning
1) Principles and Applications of Supervised Learning, Unsupervised Learning, and Reinforcement Learning
The way artificial intelligence (AI) learns on its own is broadly divided into 'Supervised Learning, Unsupervised Learning, and Reinforcement Learning.' Each learning method differs according to how data is utilized and the intended results, and performs different roles in actual industrial and technical domains where they are applied.
Understanding the process by which AI goes beyond simply processing given information to learn, adapt, and make optimal decisions on its own is a core topic in discussing AI's development. Supervised learning learns from data with correct answers given, unsupervised learning finds patterns on its own within data, and reinforcement learning acquires optimal strategies based on rewards. These learning methods are used individually, but recently hybrid models combining them to solve specific problems have emerged, contributing to building more sophisticated AI systems.
Supervised Learning: AI That Learns by Seeing Correct Answers
Supervised Learning, one of the representative ways AI learns patterns, is the process of training models based on data containing correct answers. This is similar to the process of students learning correct answers under teacher guidance, designed to acquire the ability to predict when new inputs are given.
In supervised learning, AI learns the relationship between given inputs and correct answers (outputs) and based on this predicts outputs for new data. For example, to train an AI that distinguishes dogs and cats, thousands of dog and cat images are provided and 'dog' or 'cat' labels (correct answers) are given to each image. AI learns from this data analyzing characteristics like ear shape, fur color, and size, discovering criteria for distinguishing dogs and cats on its own. After training is complete, when a new image is input, AI can predict whether the corresponding image is a dog or cat based on previously learned patterns.
This supervised learning method is utilized in various industrial fields and demonstrates powerful performance especially in areas with much structured data. In autonomous driving technology, it is used for recognizing road signs and distinguishing pedestrians and vehicles. In medical imaging analysis, it is applied to developing AI models that analyze X-ray, MRI, CT data to diagnose diseases. Voice recognition systems also utilize it — voice assistants like Apple's Siri and Amazon's Alexa are trained through supervised learning to recognize specific words and sentences and understand user commands.
Supervised learning also plays an important role in recommendation systems. Platforms like Netflix, YouTube, and Amazon use AI to analyze users' viewing histories and purchase data to recommend customized content. AI predicts which content a user is likely to prefer based on past behavior and through this provides personalized recommendation services.
As such, supervised learning is one of the most basic methods for AI to learn patterns from structured data, and when accurate data labeling is performed can derive results with very high reliability. However, limitations of supervised learning also exist. Bias included in data can be reflected directly in AI models, and continuous data updates are needed to learn new data. To solve these problems, researchers are developing methods enabling effective learning with less data, or advancing in the direction of utilizing approaches like Semi-Supervised Learning combining supervised and unsupervised learning.
Unsupervised Learning: Finding Hidden Patterns in Data
Unsupervised Learning is a method of AI analyzing given data without correct answers (labels) and exploring patterns and structures within data on its own. While supervised learning is the process of learning defined answers, unsupervised learning is closer to the process of finding hidden relationships within disorganized data. This method has differentiation in that AI learns on its own without human intervention.
In unsupervised learning, AI operates by analyzing similarity between data to form groups (Clustering) or learning relationships frequently appearing within data (Association Rule). For example, a company automatically classifying consumer groups with similar purchase propensities by analyzing customer data is a representative case of unsupervised learning. While supervised learning has humans provide correct answers, in unsupervised learning AI directly explores patterns and characteristics repeatedly appearing within data to find meaningful structures.
These characteristics of unsupervised learning are utilized in various industrial fields. In the finance and security domain, Anomaly Detection techniques can be used to detect financial fraud or network security threats, with strengths in identifying abnormal signs different from normal patterns. In marketing and customer management, through Customer Segmentation, companies analyze consumers' purchase patterns and it is used to establish customized marketing strategies for each group.
Unsupervised learning also plays an important role in natural language processing (NLP). AI can automatically extract specific topics or keywords from text data and is used to find meaningful information from vast documents like news articles, social media data, and customer reviews. Also, in medical data analysis, unsupervised learning is bringing innovative changes. It is applied in research to identify gene mutations related to specific diseases by analyzing genetic data, or to discover new disease patterns by grouping medical image data.
Unsupervised learning, possessing the ability to discover hidden patterns within data, is useful for solving complex problems that humans cannot directly define. However, unlike supervised learning, since no correct answers exist, a process of verifying whether patterns found by AI are actually meaningful is necessary. Accordingly, researchers are developing various techniques to strengthen unsupervised learning and are advancing in the direction of building more sophisticated AI models by combining with semi-supervised learning or reinforcement learning.
Reinforcement Learning: AI That Learns Based on Rewards
Reinforcement learning is a learning method where AI interacts with the environment and finds optimal actions. This is similar to the process of humans developing problem-solving ability through trial and error. AI attempts various actions to achieve specific goals, receives rewards based on results, and gradually learns methods for making better decisions.
The core principle of reinforcement learning is that AI acts as an 'Agent' selecting actions (Action) within an environment (Environment) and learning by receiving 'Rewards' according to those actions. AI continuously optimizes the strategy (Policy) maximizing rewards and finds the most advantageous actions in the long term. For example, an AI learning chess tries various moves and explores and optimizes strategies for improving winning chances on its own. Initially it moves randomly, but over time forms an increasingly sophisticated play style based on learned strategies.
Reinforcement learning is utilized in various industries requiring autonomous decision-making. A representative case is 'AlphaGo,' which utilized reinforcement learning to reach a level surpassing human professional Go players. In robotics, robots learn to pick up objects or avoid obstacles and find optimal movements. In autonomous driving systems, AI uses countless driving simulations to develop safe and efficient driving strategies. Also, in financial trading, reinforcement learning is applied where AI learns real-time market changes and automatically executes optimal investment strategies.
Reinforcement learning is contributing to solving complex problems difficult for humans to directly intervene in various fields including games, robotics, autonomous driving, and financial modeling, and is expected to be utilized in increasingly wider industrial domains going forward.
Fusion of Learning Methods: AI's Development Direction
AI's learning methods are divided into supervised learning, unsupervised learning, and reinforcement learning, but recently hybrid learning models combining these are receiving increasing attention. Approaches actively researched fuse multiple techniques to maximize advantages each learning method possesses and compensate for disadvantages.
'Semi-Supervised Learning' is a method combining supervised and unsupervised learning, training models utilizing small amounts of labeled data and large amounts of unlabeled data together. For example, in medical imaging analysis, small amounts of labeled X-ray data and large amounts of unlabeled images are learned together, enabling AI to perform more precise diagnosis. This is particularly useful in fields where data labeling costs are high and is receiving attention as an effective method to solve data shortage problems.
The combination of reinforcement learning and supervised learning is also an important factor accelerating AI's development. In autonomous driving and robotics fields, a method of first performing initial learning through supervised learning then applying reinforcement learning to find optimal actions on its own in actual environments is widely used. For example, autonomous vehicles first learn vast road driving data through supervised learning, then further sophisticate driving ability through reinforcement learning in actual road environments.
As such, AI is not dependent on a single learning method but is continuously acquiring thinking and decision-making abilities increasingly similar to humans by combining various learning techniques. AI's learning models will continue to develop going forward and will become core elements building even more sophisticated and powerful AI systems.
<Continued in next installment>
