Conversations with smart speakers, facial recognition by cameras, and generative AI that writes text——artificial intelligence has already become a tool of everyday life. However, not many people can explain “how AI becomes intelligent.” In this lecture, you will learn the basic mechanisms of machine learning, deep learning, and generative AI, as well as points to be careful about when using them.
Artificial Intelligence and the Limitations of Rule-Based Systems
Artificial intelligence (AI: Artificial Intelligence) is a general term for technologies that attempt to achieve human intellectual activities, such as recognition, reasoning, and judgment, using computers. Early AI mainly used a rule-based method, in which people entered many rules such as “if A, then do B.” Expert systems, which turned experts’ knowledge into rules, are a representative example. However, this method had the limitation that it was impossible to turn into rules recognitions that humans perform unconsciously, such as “it is impossible to write down in words all the rules for distinguishing pictures of cats.”
Machine Learning — Learning Regularities from Data
Machine learning overcame this limitation. Instead of humans writing rules, it is a method in which the computer itself finds regularities (patterns) from a large amount of data. It is classified into three types according to how learning is carried out. ①Supervised learning: The system is trained with data that has correct labels (for example, images marked “This is a cat” and “This is a dog”), and it classifies unknown data or predicts numerical values. ②Unsupervised learning: Without being given correct answers, the system is made to find groups of similar items (clusters) or structures in the data. It is used for grouping customers and similar tasks. ③Reinforcement learning: Instead of correct answers, the system is given “rewards” and learns actions that maximize rewards through trial and error. It is used in game AI and robot control.
Deep Learning
Among machine-learning methods, deep learning has driven the recent AI boom. It is a neural network that mathematically models connections among nerve cells (neurons) in the human brain and is deeply stacked into many layers. In conventional machine learning, humans had to design “where in an image to pay attention” (features), but deep learning was groundbreaking because it can automatically acquire the features themselves from data. From around 2012, the accuracy of image recognition improved dramatically, and systems achieved performance equal to or better than humans in speech recognition, automatic translation, and Go AI. The success of deep learning became possible because three conditions came together: ① a large amount of data (big data), ② improved computing power (the use of GPUs), and ③ improved algorithms.
The Emergence of Generative AI
In the 2020s, generative AI, which creates new text, images, speech, programs, and other content, spread rapidly. At the core of text-generating AI are large language models (LLMs), which are trained to predict “the word that is most likely to come after a sequence of words” from enormous amounts of text on the Internet. Although they can fluently create plausible sentences, because of how they work, they may produce content that is not factual with an air of confidence (hallucination). The basic premise for using generative AI is to understand that its output is “a statistically natural continuation,” not “a verified fact.”
Points to Note When Using AI
When using AI, pay attention to the following points. ①Bias in training data directly becomes bias in judgments (for example, AI trained on past hiring data may unfairly rate people with certain attributes lower). ② Deep learning tends to become a black box whose basis for judgment is difficult for humans to explain, and research on explainable AI (XAI) is progressing. ③ The user and humans are responsible for verifying the output and making the final decision. AI is a powerful tool, but knowing and using its “quirks” is the core of learning Information II.
- Rule-based = humans write the rules. Machine learning = the computer learns regularities from data.
- Three types of machine learning: supervised learning (classification and prediction using labeled data), unsupervised learning (discovering structures), and reinforcement learning (trial and error using rewards).
- Deep learning = multilayer neural networks. Its groundbreaking feature is that it can acquire features automatically.
- Three conditions for the success of deep learning: big data, computing power (GPUs), and improved algorithms.
- Generative AI (LLMs) creates text by predicting “the word most likely to come next.” Be careful of hallucinations.
- Understand bias and black-box problems; humans are responsible for verification and final decisions.
Practice Questions
- Explain the difference between rule-based AI and machine learning, focusing on “who creates the rules.”
- Explain the differences among “supervised learning,” “unsupervised learning,” and “reinforcement learning” so that it is clear how each system is trained.
- Using the term “features,” explain what was groundbreaking about deep learning compared with conventional machine learning.
- Explain what “hallucination” in generative AI is, and state why it occurs based on how large language models work.
Answers and Explanations
- Answer: Rule-based AI is a method in which humans enter judgment rules such as “if A, then B” in advance. Machine learning is a method in which the computer itself finds regularities (patterns) from a large amount of data instead of humans writing the rules. Recognitions that cannot be fully described in words, such as distinguishing images, became possible through learning from examples.
Explanation: The key is the contrast between “humans write the rules” and “machines learn from data.” A complete answer should also mention the significance of becoming able to solve problems that cannot be turned into rules. - Answer: Supervised learning is a method in which the system is trained with data labeled with correct answers and classifies or predicts unknown data. Unsupervised learning is a method in which the system finds groups of similar items or structures in data without being given correct answers. Reinforcement learning is a method in which the system is given rewards instead of correct answers and learns actions that maximize rewards through trial and error.
Explanation: Organize the differences according to the three learning signals: “with correct answers / without correct answers / with rewards.” Understanding is deeper when specific examples such as image classification, customer grouping, and game AI are included. - Answer: In conventional machine learning, humans had to design features that indicate where in the data the system should pay attention. Deep learning uses multilayer neural networks to automatically acquire the features themselves from a large amount of data, thereby achieving dramatic improvements in recognizing complex patterns, such as images and speech, that humans cannot express in words.
Explanation: “Automatic acquisition of features” is the core of the answer. It is also good to mention the structure of multilayer neural networks. - Answer: Hallucination is a phenomenon in which generative AI outputs content that is not factual in a plausible and confident manner. Large language models are trained to predict “the word most likely to come after a sequence of words,” so their output is statistically natural text, not text that has been verified for factual accuracy. As a result, they may fluently generate information that does not actually exist.
Explanation: In addition to explaining the phenomenon, the main point for grading is whether the answer states the reason that accuracy is not guaranteed because of the next-word prediction mechanism. Countermeasures include checking sources and cross-checking.