Recommended
More Related Content
Similar to 1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf 1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf
Similar to 1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf 1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf (20)
1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf 1017_3_ChatGPT挑戰解方及商機_陳縕儂.pdf
- 1. ChatGPT帶來的挑戰、 解方與未來商業機會 陳縕儂 Yun-Nung (Vivian) Chen http://vivianchen.idv.tw
- 2. 陳縕儂 Yun-Nung (Vivian) Chen 國立台灣大學資訊工程學系 副教授 美國微軟研究院 博士後研究員 美國卡內基美隆大學 碩士 / 博士 國立台灣大學資訊工程學 學士 / 碩士 台灣傑出女科學家新秀獎 吳大猷先生紀念獎 傑出人才基金會年輕學者創新獎 Amazon AWS ML Research Awards Google Faulty Research Awards
- 3. AI & ML Artificial intelligence (AI) is intelligence—perceiving, synthesizing, and inferring information—demonstrated by machines, as opposed to intelligence displayed by animals and humans. Machine learning (ML) is a field of inquiry devoted to understanding and building methods that “learn”, that is, methods that leverage data to improve performance on some set of tasks. It is seen as a part of artificial intelligence.
- 4. • Speech Recognition • Handwritten Recognition • Weather forecast Learning ≈ Looking for a Function f f f “2” “你好” “ Tuesday” Sunday
- 5. Deep Neural Network • Cascading the neurons to form a neural network 5 1 x 2 x … … Layer 1 … … 1 y 2 y … … Layer 2 … … Layer L … … … … … Input Output M y N x Each layer is a simple function in the production line
- 6. History of Deep Learning • 1960s: Perceptron (single layer neural network) • 1969: Perceptron has limitation • 1980s: Multi-layer perceptron • 1986: Backpropagation • 1989: 1 hidden layer is “good enough”, why deep? • 2006: RBM initialization (breakthrough) • 2009: GPU • 2010: breakthrough in Speech Recognition (Dahl et al., 2010) • 2012: breakthrough in ImageNet (Krizhevsky et al. 2012) • 2015: “superhuman” results in Image and Speech Recognition • 2016: AlphaGo “superhuman” results in Go playing • 2022: ChatGPT “human-level” results in diverse domains 6 Why does deep learning show breakthrough in applications after 2010?
- 7. Why Deep Learning Works Big Data GPU
- 8. GPU for Efficient AI Training
- 9. Difference from Prior AI 近期生成式AI與過去的差異 9
- 10. Search Engines for Information Seeking
- 11. When we don’t know what to search
- 12. Extractive 抽取式 vs. Generative 生成式
- 13. 微軟小冰 vs. ChatGPT • Extractive-based o can only answer questions similar to seen ones o answers may be more reliable o consider short context in a conversation • Generative-based o can answer any unseen questions o answers needs to be validated o consider whole context in a conversation
- 14. Beginners vs. Professionals • Extractive-based o can only answer questions similar to seen ones o easy to control • Generative-based o can answer any unseen questions based on the known domain knowledge o difficult to control
- 15. Learning Pyramid ● Bloom’s Taxonomy (2001) CREATE EVALUATE ANALYZE APPLY UNDERSTAND REMEMBER Use information to create some thing new Grasp meaning of instructional materials Take apart the known and identify relationships Use information in a new (but similar) situation Examine information and make judgements Recall specific facts High-order thinking skills Low-order thinking skills
- 16. GPT: Generative Pretrained Transformer (Radford et al., 2018) • Pre-trained on BooksCorpus (~7000 books; 5GB) 16 Transformer Vivian goes to make tasty tea goes to make tasty tea Radford, Alec, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. "Improving language understanding by generative pre-training." (2018). <EOS> <BOS> Vivian Vivian goes to make tasty tea 文字接龍
- 17. 17 GPT-3 Prompting • Zero-Shot • One-Shot • Few-Shot
- 18. 18 OpenAI GPT Paradigm Pre-Trained Data #Parameters Model 5GB 0.117 B GPT (Radford et al., 2018) 40GB 1.5 B GPT-2 (Radford et al., 2019) 45TB 175 B GPT-3 (Brown et al., 2020) ? ? GPT-4 (OpenAI, 2023)
- 19. Emergence 湧現 Scale of language model
- 20. AI Objectives Helpful 有用 help the user solve their task Honest 誠實 shouldn’t fabricate information or mislead users Harmless 無傷害性 should not cause physical, psychological, or social harm to people or the environment
- 21. Helpful Current output What we want
- 22. Honest • Hallucinations: output that cannot be verified from the source content o Pros: factual hallucination improves informativeness of the generated text o Cons: unverifiable information increases the risk from a safety perspective 22 Source Inform (name=pickwick hotel, pricerange=moderate) Output the hotel named pickwick hotel in san diego is in a moderate price range
- 24. Issues in Pre-Trained Models 1. Helpful – help the user solve their task Unable to follow the user’s instructions 2. Honest – shouldn’t fabricate information or mislead the user Hallucinations 3. Harmless – should not cause physical, psychological, or social harm to people or the environment Not easy to detect/identify 24 Solution: RLHF (Reinforcement Learning from Human Feedback)
- 25. ChatGPT: Reinforcement Learning from Human Feedback • Improving GPT via teacher’s feedback generation update via reinforcement learning Teacher-given scores a conversation history a model-generated output : 我想了解台灣的知名女歌手。 : 台灣知名女歌手包含蔡依林、… : 可以替我介紹㇐下蔡依林嗎？ 當然沒問題！蔡依林曾獲x次金曲獎… Human Feedback
- 28. Good at: Open Questions
- 29. Good at: Open Questions
- 30. Good at: Text Drafting
- 31. Good at: Text Drafting
- 32. Edge Browser with AI Copilot
- 33. Limitations: Training Data Bias
- 34. Limitations: Training Data Bias
- 38. 資訊有錯誤怎麼辦？ 1. “Human-in-the-loop” 人類把關 • 專家確認 AI 的回答是否正確; 適當追問、修改 • 人類承擔此責任
- 39. 資訊有錯誤怎麼辦？ 2. Domain-specific pre-training/fine-tuning • 加強特定領域知識 rarely changed domain-specific data pre-training / fine-tuning : :
- 40. 資訊有錯誤怎麼辦？ 3. Grounding on (dynamically changed) knowledge resources • 先進行搜尋 • 搜尋到相關資訊後再看著資料回答 dynamically changed reference data
- 41. 顧客資料外流怎麼辦？ 1. Microsoft Partnership • 模型使用授權 • 私人資料保護 fast development, better performance
- 42. 顧客資料外流怎麼辦？ 2. 建置自己的模型 • 訓練自家的 ChatGPT • 合成資料以增進自家原本 AI 模型效能 safer, lower cost Alpaca (Stanford) LLaMA (Meta) Vicuna (Berkeley, UCSD, CMU)
- 43. Taiwan-LLaMa: LMs for Taiwanese Culture https://github.com/MiuLab/taiwan-llama
- 44. Taiwan-LLaMa Demo Taiwan LLaMa v1.0 ChatGPT Question 高雄以擁有惡地景觀著名的景點是田寮 月世界。✅ 以惡地地景著名的是高雄市的旗津。 以惡地景觀著名的是 高雄哪㇐個景點? 唐鳳是政治評論家、資訊社會學研究者， 也是台灣第㇐位擔任政府數位政委的 人。 ✅ 唐鳳（Audrey Tang）是台灣的政治家、 社會活動家和軟體工程師。她於2016年 擔任台灣行政院數位政務委員，是台灣 政府歷史上首位沒有政黨背景的政務官。 唐鳳在軟體工程領域有豐富的經驗。 ✅ 唐鳳是誰？ NTU 是國立臺灣大學的縮寫。它是臺灣 的㇐所頂尖研究型大學，位於臺北 市。 ✅ NTU位於新加坡，具體地址是50 Nanyang Avenue, Singapore 639798。 NTU 在哪? "22K"是指台灣的大學畢業生的起薪，大 約是新台幣22,000元。這個詞被用來形 容大學畢業生的起薪偏低，並且在就業 市場上對於大學畢業生的需求不足。✅ "22K" 可能指的是金的純度，尤其是在珠 寶和金飾方面。 甚麼是22K?
- 47. 為何現在開始？ Figure from GPT-4 Technical Report 十萬倍
- 50. 54% 5% 5% 5% 1% 30% 英文 西班牙文 俄文 德文 中文 其他 Numbers from https://en.wikipedia.org/wiki/Languages_used_on_the_Internet 課本從哪來？
- 52. 背誦課文 • 從 LLaMA2 (13B) 開始進行預訓練 • 維基百科 • 網路資料 (CommonCrawl) # Tokens Category Datasets 427,662,690 Written (Knowledge) Wikipedia-zht 4,971,287,180 General (Web) CC100-zht
- 53. 背誦課文 • ⾧度上限: 四千 • 每㇐個文本只看過㇐次 (1 Epoch) • 八張 Nvidia A100 訓練兩週 • 加速技巧 Fully Sharded Data Parallel Flash Attention v2 Bfloat16 降低記憶體用量 降低運算精準度
- 54. 背誦課文
- 58. 學以致用 • 涵蓋了 15 組模擬考題 • ~ 五十萬條對話
- 59. 學以致用 • 從 Taiwan-LLaMa-Base 開始進行訓練 • ⾧度上限: 四千 • 每㇐個範例看過三次 • 八張 H100 訓練了十二小時
- 62. 25% 75% ChatGPT 錯誤 正確 25% 75% Taiwan LLaMa 錯誤 正確 Question from https://arxiv.org/pdf/2303.04715.pdf 台灣大小事問答集
- 63. 如何微調成企業級應用?
- 65. Data Bias Issues
- 66. Data Bias Issues
- 71. ChatGPT Plugins
- 77. Responsible AI Transparency and Accountability: transparency in how AI systems make decisions and operate; accountable for the outcomes o 《歐盟人工智慧法》將要求公司們，對於這些模型多加說明。如果公司想在歐盟銷售或使用 AI 產 品，但不遵守這些透明化的義務，可能就會面臨最高全球總營業額 6% 的罰款。 Ethical and Fair Use: ensuring fairness by reducing biases in AI systems to avoid discrimination Privacy and Safety: privacy protection and robustness against attacks and harm The EU wants to regulate your favorite AI tools | MIT Technology Review
- 78. Elon Musk open-sourced Twitter’s algo
- 79. 法律規範 經濟部智慧財產局 電子郵件1111031 民國 111 年 10 月 31 日 要旨： ㇐、依我國著作權法第3條第2款及第10條規定，著作人指創作著作之人，著作人於著作完成時 享有著作權。換言之，著作必須係以自然人或法人為權利義務主體的情形下，由自然人所為的 創作，方可能受到著作權的保護。 二、有關人工智慧(AI)的創作，大致上可分為兩種： (㇐)第㇐種是「以人工智慧為工具的創作」，也就是人類有實際的創意投入，只是把人工智慧 (例如：繪圖軟體)當作輔助工具來使用，在這種情形依輔助工具投入創作者的創意而完成的創 作成果仍可以受著作權保護，著作權則由該投入創意的自然人享有，除非有著作權法第11條及 第12條之情形。 (二)第二種是「人工智慧獨立創作」，也就是人類並無實際的創意投入，完全是由AI的演算功 能獨立進行完成創作，此時由於AI並非自然人，沒有人類精神文明的投入，其創作完成成果自 然不屬於著作權法保護的著作，原則上無法享有著作權。
- 80. Real or Fake? • Google helps identify if the image is generated by AI 1) Show where on the web it first appeared and which sites it’s appeared on
- 81. • Google helps identify if the image is generated by AI 1) Show where on the web it first appeared and which sites it’s appeared on 2) Self-labeling in meta-data Real or Fake?
- 82. Data Bias Issues
- 83. Data Bias Issues
- 86. Impact on Human’s Mentality
- 87. CREDITS: This presentation template was created by Slidesgo, and includes icons by Flaticon, and infographics & images by Freepik