AI Ch.01 - HackMD

--- title: AI Ch.01 --- # Artificial Intelligence NTNU 人工智慧 ##### [Back to Note Overview](https://reurl.cc/XXeYaE) ##### [Back to Artificial Intelligence](https://hackmd.io/@NTNUCSIE112/AI110-2) {%hackmd @sophie8909/pink_theme %} ###### tags: `AI` `110-2` `CSIE` `選修` `NTNU`   ## Ch.0 ### 什麼是智慧？ > Intelligence: The ability to learn and solve problems [name=Webster's Dictionary] - "Philosophia"：愛好智慧 - 1985 年美國心理學家 Sternberg 提出智慧三元論。 - Emphasizes how 3 types of abilities work together to create intelligent behavior - Triarchic Theory - Analytical - compare, evaluate and analyze - Creative - insights, synthesis, adapting in unique situations - Practical - dealing with everyday tasks relating to world ### AI 是否會主宰人類世界 > 我們很有必要給人工智慧加上一個「紅色按鈕」，用來阻止其傷害人類和環境...，但人翁智慧也有可能學會防止自己被人類關機...[name=Deepmind和牛津大學] > 人工智慧終究會帶來人類的末日 [name=霍金] > AI 是人類生存最大的威脅！AI 可能比核武器更危險[name=特斯拉創辦人馬斯克] ### AI 的兩個重要里程碑 - 1997年：IMB 深藍（Deep Blue）擊敗西洋棋棋王 Kasparov - 2016年：Google Deepmind's AlphaGp 擊敗李世石（Lee Sedol） - 2017 3 場完勝柯潔 ### 強弱 AI - 弱人工智慧：認為**不可能**製造出能真正地推理和解決問題的智能機器，只是看起來像是智能的但不是真正擁有智能，也不會有自主意識 - 強人工智慧：認為**有可能**製造出能真正地推理和解決問題的智能機器，這樣的機器被認為是有知覺和有自我意識的。強人工智慧有分兩種 - 類人的人工智慧：機器的思考和推理就像人的思維一樣 - 非類人的人工智慧：機器產生了和人完全不一樣的知覺和意識，使用和人完全不一樣的推理方式 ## Ch.01 Introduction ### 1.1 What is AI? #### AI 定義 - Systems that ... | | fidelity 保真性 | rationality 理性 | | -----------------: | :-----------------: | :----------------: | | thought/reasoning | think like humans | think rationally | | behavior | act like humans | act rationally | - Acting humanly - Turing test - The test is for a program to have a conversation via online typed messages with an interrogator who has to guess if the conversation is with a machine or a person - The program passes the test if it fools the interrogator（審訊者） - Total Turing Test - require interaction with objects and people in real world - Acting rationally - Rational Agent - just something that perceives and acts - autonomously, adaptively, goal-directly - do the right thing - expected to maximize the goal achievement, given the available information - **don't necessarily** involve think/inference - e.g. car factory - Thinking Humanly - cognitive modeling 認知模型 - 對人（或動物）認知過程的近似 - require - to **know how** people think - scientific theories of internal activities of the brain - once have a sufficiency precise theory of human mind, we can express it as a computer program - Thinking Rationally - Laws of thought - correct inference - correct premises yield correct conclusions - formal logic - define a **precise notion** for statements about all things and the relations among them - main considerations - not all things can be represented in logic forms - computation complexity is high - example: 四色定理、布林畢氏三元數問題、證明 n = 7825 不能成功著色 ### 1.2 Foundations of AI - Philosophy 哲學 (428 B.C. - present) - **logic, methods of reasoning** - Aristotle（亞里斯多德）：A set of rules that can describe the formal/rational part of mid. - Mind as a physical system/computation process - Knowledge acquired from experiences and encoded in mind, and used to choose right actions. - Learning, language, rationality - Mathematics 數學 (800 - present) -** Formal representation and proof** - Tools to manipulate logical/probabilistic statements - Groundwork for computation and algorithms - Three main contributions: - logic, computation, and probability - Economics 經濟學 (1776 - present) - Formal theory for making decisions - Utility: the preferred outcomes - Decision theory - Game theory - Operations research - Neuroscience 神經科學 (1861 - present) - Brain cause minds: The mapping between areas of the brain and the parts of body they control or from which they receive sensory input - Ramon y Cajal（拉蒙卡哈）：神經解剖學者、1906 諾貝爾獎得主 - 提出新學說：腦神經是由個別的神經元組成，在神經內訊號以單一方向傳遞訊號（樹突→細胞核→軸突→突觸）。 - 人腦約有 500-1000億個（$10^{11}$）神經元，速度 $10^{-3}$ sec - 矽晶片 $10^9$ sec - Psychology 物理學 (1879 - present) - Brains as information-processing devices - Knowledge-based agent - Stimulus（刺激） translated into an internal representation - Cognitive process derive new representation from it - These representations are in turn retranslated into action - Computer engineering 計算機工程 (1940 - present) - Artifacts for implementing AI ideas - (Software) programming languages - The increase in speed and memory - Control theory 控制理論 (1948 - present) - Design simple optimal agents receiving feedback from the environment - Maximizing an objective function over time - Minimize the difference between current and goal states - Linguistics 語言學 (1957 - present) - How does language relate to thought? - Languages fit information processing model - Understanding languages requires an understanding of subject matter and context ### 1.3 History of AI - 1943-55 Gestation（孕育） of AI - McCulloch & Pitt : Boolean circut model of neurons - Turing's "Computing Machinery and Intelligence" - **1956 The birth of AI** - Darthmouth meeting: "Artificial Intelligence" adopted - 1966 ELIZA: a simple natural-language understanding program - 1966-85 Neural Network almost disappear - No efficient training algorithm for layered networks - 1969-79 Knowledge-based systems - 1968 MIT Microworlds: "(geometric) analogy problem" - 1972 MIT Microworlds: "blocks world" SHRDLU, rearrange the blocks in a certain way - 1980-88 Expert system industry booms - a million dollars to billion of dollars - 1981 Japanese's 5th generation Computer running Prolog - 1986- Neural networks return to popularity - 1988-93 Expert system busts："AI winter" - 1996 IBM Deep Blue 打輸世界西洋棋棋王 2-4 - 1997 IBM Deep Blue 打敗世界西洋棋棋王 **Garry Kasparov** 3.5-2.5 - 2006- Marvin Minsky: The Emotion Machine - 2001-present Large datasets. Data will drive future discoveries and alleviate the complexity in AI - 2001 Banko and Brill, word-sense disambiguation（歧義解釋） - 2007 Hays and Efros , filling in holes in a photograph - 2010-present Machine learning is everywhere - 2016 03, `AlphaGo` defeated the Go﹙圍棋﹚ champion **Lee Sedol**, a 9 dan profession, 4-1. ### 講義填空 - Ch1-21 `AlphaGo` - ### Ch1後實作類神降算神秘函數 #### Code :arrow_right: [repl.it](https://replit.com/languages/python3) :::spoiler sample code ```python= #講義第六頁 sample code from numpy import * train_in = array([[0,2,0],[0,0,1],[0,1,1],[1,0,1],[1,1,1]]) train_sol = array([[0,0,0,1,1]]).T # Initialize nn_weights random.seed(1) nn_weights = 2 * random.random((3, 1)) - 1 # Train the network for i in range(200): print("\n i= ", i, "nn_weight=") print(nn_weights) # Calculate the outputs for each training examples train_out = 1 / (1 + exp(-(dot(train_in, nn_weights)))) print("train_out =") print(train_out) # Run the NN adjustments nn_weights += dot(train_in.T, (train_sol-train_out)*train_out*(1-train_out)) # Unknown test input test_in = array([1,0,0]) # Print the result for our unknown test input print('\nThe final prediction is ',1/(1 + exp(-(dot(test_in, nn_weights))))) ``` ::: #### 神經元 ![](https://i.imgur.com/voGPIKG.png) cell body sigma wixi+b