AI 魔法實驗室
嚴謹的課程結構整合了四大主要部分：人工智能基礎、大型模型生成（GenAI 與 LLM）、智能體與演化計算（以香港理工大學特色為重點），以及倫理。課程邏輯按順序推進：感知與數據（第1-3課），認知與生成（第4-6課），智能體與演化（第7-9課），最後以倫理與未來（第10課）作結。
課程總覽
📚 內容概要
「AI魔法實驗室」是一門嚴謹且整合性的課程，旨在提供對現代人工智慧的深入、循序漸進的理解。課程內容分為四個逐步遞進的模塊：基礎概念（感知與資料）、先進生成能力（透過大型語言模型與擴散模型的認知與生成）、自主系統（代理與演化計算），以及倫理治理（倫理與未來）。學生將從理解資料的原始數值表達出發，逐步掌握複雜系統設計，最終形成對負責任的人工智慧創造與部署的全面觀點。
本課程提供對現代人工智慧的嚴謹且整合性的理解，涵蓋核心資料基礎知識、大型語言模型（LLM）的生成技術、自主代理的架構，以及實現負責任部署所必需的重要倫理考量。
🎯 學習目標
- 掌握人工智慧感知、資料表示（張量）及基礎監督式學習任務（如分類）的根本原理。
- 理解並掌控大型語言模型（LLM）與生成式人工智能，透過序列預測、注意力機制，以及高階提示工程技術加以應用。
- 設計並分析智能代理，結合感知-決策-行動迴圈，並整合先進的群體優化方法，例如演化計算。
- 区分生成式與判別式人工智能，並解釋擴散模型在文字轉圖像生成中的機械過程。
- 評估當代人工智慧固有的倫理挑戰（資料偏見、模型幻覺、深度偽造），並提出促進人機共生的策略。
課程 共 10 课时 · 预计 20.0h
課程
Lesson
This lesson introduces the fundamentals of Computer Vision by explaining how machines interpret images as grids of numerical pixel data. Students will learn how computers use RGB color models and feature extraction to identify shapes, edges, and patterns within digital images.
This lesson explores how computers use encoding to translate non-structured data, like images and sounds, into numerical formats that machines can process. Students learn that vectors and tensors serve as the essential mathematical containers that organize this data, allowing AI models to identify patterns and learn effectively.
This lesson introduces supervised learning, where humans provide labeled training data to help AI identify patterns and establish decision boundaries for classification. Students learn how to evaluate these models by calculating accuracy and avoiding the pitfalls of overfitting.
This lesson explores how AI processes language by breaking text into numerical tokens and managing memory through a limited context window. Students will also learn how the Attention Mechanism acts as a digital spotlight, allowing the AI to identify and focus on the most relevant words to understand complex meanings.
This lesson introduces prompt engineering as a powerful skill that allows you to use natural language to command AI, which functions as a sophisticated prediction engine. By mastering the "Prompt Blueprint"—consisting of instructions, context, and output—along with techniques like personas and few-shot learning, you can transform from a passive user into an architect of information.
This lesson explores the difference between discriminative AI, which sorts existing data, and generative AI, which creates new content from scratch. Students learn how generative models use a process called diffusion to transform random, high-entropy noise into clear, detailed images through iterative de-noising.
This lesson introduces the Perception-Decision-Action (PDA) loop, the continuous cycle that allows intelligent agents to observe their environment, plan goals, and take action. Students will also learn how sensors and actuators serve as the essential bridge between an agent's digital logic and the physical or virtual world.
This lesson introduces Evolutionary Computation, a method that mimics natural selection to solve complex problems by iteratively improving digital solutions. Students learn how to evolve AI by testing the performance of "phenotypes" and using genetic processes like mutation and crossover to refine their underlying "genotype" code.
This lesson explores how multi-agent systems use communication, collaboration, and competition to solve complex problems more effectively than a single AI. Students learn how decentralized "swarm intelligence" allows groups of agents to achieve emergent, coordinated behavior by following simple local rules.
This lesson explores the ethical challenges of AI, including algorithmic bias, hallucinations, and the risks of intentional deception through deepfakes. Students learn the importance of digital literacy and the "Human-in-the-Loop" approach, which emphasizes that humans must remain the final supervisors to ensure technology is used fairly and accurately.