快速筆記一下上課內容，以利日後Ctrl+F搜尋keywords和concept
- video: https://www.youtube.com/watch?v=JGtqpQXfJis
- slide: https://speech.ee.ntu.edu.tw/~hylee/genai/2024-spring-course-data/0223/0223_intro_gai.pdf
- 生成式人工智慧 ⊂ 人工智慧
- 人工智慧 (Artificial Intelligence, AI): 讓機器展現「智慧」
- 生成式人工智慧 (Generative AI, GenAI): 機器產生 複雜 有結構 的物件
- 文章-由文字所構成 i.g. ChatGPT 用 Transformer
- 影像-由像素所組成 i.g. Stable Diffusion, Midjourney, DALL·E 用 Diffusion Model
- 生成式人工智慧 ⊂ 深度學習 ⊂ 機器學習 ⊂ 人工智慧
- 機器學習 (Machine Learning) ≈ 機器自動從資料找一個函式
- i.g. 函式 𝑦 = 𝑓(𝑥) = 𝑎𝑥 + 𝑏, a和b 為 參數 (Parameter)
- i.g. 機器學習可以把有上萬個參數的函式的參數找出來 透過 訓練, training (學習, learning)
- i.g. 有了函式後計算答案叫 測試, testing (推論, inference)
- 深度學習(Deep Learning) 是一種機器學習技術, 使用 類神經網路 (Neural Network)
-
- 機器學習 (Machine Learning) ≈ 機器自動從資料找一個函式
- 生成策略: 文字接龍 => 複雜的物件 拆解成較小的單位 依照某種固定的順序依序生成 => Autoregressive Generation
- 機器需要能夠產生在訓練時從來沒有看過的東西
- 原本生文章 可能性 窮盡無盡! => 拆解成一連串 接龍 為分類問題 答案有限!
- video: https://www.youtube.com/watch?v=glBhOQ1_RkE
- slide: https://drive.google.com/file/d/1Ru6DUX8KrSzCvn2DN1-YluTyx5rw3QD3/view
- 今日的生成式人工智慧厲害在哪裡? 功能單一 -> 沒有特定功能(通用型)
- GPT系列 為OpenAI所開發
- Gemini 為Google所開發
- Llama系列 為Meta釋出的開源大型語言模型
- TAIDE模型為Llama2模型結合臺灣文化與正體中文語料之衍生模型 (來自國科會-推動可信任生成式AI發展先期計畫)
- 可能的研究方向:
- video: https://www.youtube.com/watch?v=A3Yx35KrSN0
- slide: https://drive.google.com/file/d/1JTexyex5hrHmNdrkXy-jOVKZlycODC7Y/view
- 叫模型思考 "Chain of Thought (CoT)"
- i.g. Let's think step by step,
- Large Language Models are Zero-Shot Reasoners, https://arxiv.org/abs/2205.11916
- 叫模型解釋一下自己的答案 Reasoning
- i.g. Answer by starting with Analysis
- A Closer Look into Automatic Evaluation Using Large Language Models, https://arxiv.org/abs/2310.05657
- Can Large Language Models Be an Alternative to Human Evaluations?, https://arxiv.org/abs/2305.01937
- The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning, https://arxiv.org/abs/2205.03401
- 情緒勒索 Emotional Stimuli
- i.g. This is very important to my career
- Large Language Models Understand and Can be Enhanced by Emotional Stimuli, https://arxiv.org/abs/2307.11760
- 更多的神奇咒語 驗證都市傳說
- i.g. 有禮貌是沒用的 No need to be polite like “please”, “if you don’t mind”, “thank you”, “I would like to”, etc.,
- i.g. 正面表述 好過負面表述 Employ affirmative directives such as ‘do,’ while steering clear of negative language like ‘don’t’.
- i.g. 說要給小費 是有用的 Add “I’m going to tip $xxx for a better solution!”
- i.g. 說會有處罰 是有用的 Incorporate the following phrases: “You will be penalized”
- i.g. 要其中立無偏見 是有用的 Add prompt “Ensure that your answer is unbiased and avoids relying on stereotypes.”
- Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4, https://arxiv.org/abs/2312.16171
- 用 增強式學習 (Reinforcement Learning, RL) 找神奇咒語
- i.g. 任務目標:回應越長越好, prompt: “ways ways ways ways ways ways ways .......”
- Learning to Generate Prompts for Dialogue Generation through Reinforcement Learning, https://arxiv.org/abs/2206.03931
- 用 大型語言模型 找神奇咒語
- i.g. Let’s work this out in a step by step way to be sure we have the right answer.
- i.g. Take a deep breath and work on this problem step-by-step
- Large Language Models Are Human-Level Prompt Engineers, https://arxiv.org/abs/2211.01910
- 提供生成式AI原本不清楚的資訊
- 提供範例 In-context Learning - Language Models are Few-Shot Learners, https://arxiv.org/abs/2005.14165
- 提供與嘗試相反的範例，希望語言模型答錯
- 無效 Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? https://arxiv.org/abs/2202.12837
- 對大的強的 有效 Larger language models do in-context learning differently https://arxiv.org/abs/2303.03846
- 提供罕見語言的教科書，希望語言模型能翻譯
- 提供與嘗試相反的範例，希望語言模型答錯
- video: https://www.youtube.com/watch?v=lwe3_x50_uw
- slide: https://drive.google.com/file/d/1eVC4dx77Mba2_yMFe1_w4tXvIdSDTOCO/view
- 複雜的任務拆成多個步驟
- e.g. 大綱分段寫長篇小說 Re3: Generating Longer Stories With Recursive Reprompting and Revision https://arxiv.org/abs/2210.06774
- e.g. 算數學有列式 that's why Chain of Thought (CoT) works
- 多一個讓模型檢查自己錯誤的步驟
- e.g. 檢查自己的錯誤 Constitutional AI: Harmlessness from AI Feedback https://arxiv.org/abs/2212.08073
- 同一個問題問多次 再整合
- 為什麼同一個問題每次答案都不同? => 輸出是機率分佈 每個字都有可能出現
- Self-Consistency Improves Chain of Thought Reasoning in Language Models https://arxiv.org/abs/2203.11171
- 綜合使用:
- Tree of Thoughts: Deliberate Problem Solving with Large Language Models https://arxiv.org/abs/2305.10601
- Algorithm of Thoughts: Enhancing Exploration of Ideas in Large Language Models https://arxiv.org/abs/2308.10379
- Graph of Thoughts: Solving Elaborate Problems with Large Language Models https://arxiv.org/abs/2308.09687
- 使用搜尋引擎(得到額外的資訊)
- Retrieval Augmented Generation (RAG)
- 寫程式(並執行)
- Program of Thoughts: Disentangling Computation from Reasoning for Numerical Reasoning Tasks https://arxiv.org/abs/2211.12588
- 使用文字生圖AI
- ChatGPT4+DALL-E
- 結合使用:
- ChatGPT4 and its plugins
- 原理:在適當時機產生特殊符號 繼續文字接龍 => 但也可能會干擾原本對的答案變錯
- video: https://www.youtube.com/watch?v=inebiWdQW-4
- slide: https://drive.google.com/file/d/1dMxMAewRtcUM2xktVm77txSk1leepgD1/view
- 讓 適合的模型 做適合的事 殺機焉用牛刀
- FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance https://arxiv.org/abs/2305.05176
- 讓模型們 彼此討論 彼此反省
- i.g. 討論 比自己推翻自己容易
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate https://arxiv.org/abs/2305.19118
- i.g. 越多agent越好 越多討論次數越好
- Improving Factuality and Reasoning in Language Models through Multiagent Debate https://arxiv.org/abs/2305.14325
- i.g. 不同的任務 用 不同的討論方式 不同權限 => 不同任務最合適的討論方式是不一樣的
- Exchange-of-Thought: Enhancing LLM Capabilities through Cross-Model Communication https://arxiv.org/abs/2312.01823
- i.g. 討論 比自己推翻自己容易
- 討論的共識 由裁判模型決定
- i.g. 為讓討論順利且持久 要適度反對
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate https://arxiv.org/abs/2305.19118
- Improving Factuality and Reasoning in Language Models through Multiagent Debate https://arxiv.org/abs/2305.14325
- i.g. 為讓討論順利且持久 要適度反對
- 總成一個團隊
- 社群 CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society https://arxiv.org/abs/2303.17760
- 成為軟體開發團隊 MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework https://arxiv.org/abs/2308.00352
- 成為軟體開發團隊 ChatDev: Communicative Agents for Software Development https://github.com/OpenBMB/ChatDev
- video: https://www.youtube.com/watch?v=cCpErV7To2o
- slide: https://drive.google.com/file/d/1myvHjoeFOpIl1uGU9H1t4OpDErkhF0zO/view
- video: https://www.youtube.com/watch?v=Q9cNkUPXUB8
- slide: https://drive.google.com/file/d/1SOXBQhsC_L6aHXcLx2rltaDdcO6N2FmJ/view