Ai plus-ai intro 02-20170605
2 likes338 views
The document provides a comprehensive overview of artificial intelligence, tracing its history from Ada Lovelace in 1842 to significant developments in deep learning and machine learning. It discusses key milestones in AI research, notable projects like AlphaGo, and the evolution of AI technologies, particularly in the context of autonomous vehicles in China. Additionally, it references various academic sources and AI principles aimed at benefiting society.
Ai plus-ai intro 02-20170605
- 4. 望梅止渴
- 11. Src: Fortune, 2016/09 Training with 2 hidden layers Training with 4 hidden layers
- 21. AI Timeline Ada (1842) Alan Turing (1950) The first conference on AI by John McCarthy, Marvin Minsky (1956) Demonstrated by Newell (1957) Unimation s working on GE (1961) Joseph Weizenba um (1965), E. Geigenba um (1965) Chess- playing program by Greenblatt at MIT (1968) Jack Myers Harry Pople (1979) 1980s Ian Horswil l (1993) TiVo Suggestions (2005) Apple, Google, Micorsoft (2011) Machine Learning, Deep Learning (2013 ~)
- 22. KEY MOMENTS IN DEEP- LEARNING HISTORY
- 23. KEY MOMENTS IN DEEP- LEARNING HISTORY 擊碎玻璃天花板的華裔女科學家 http://deeplearning.net/2012/12/13/googles-large-scale-deep-learning-experiments/
- 24. KEY MOMENTS IN DEEP- LEARNING HISTORY
- 29. Deep Dream
- 30. Deep Dream
- 31. Deep Dream
- 32. Deep Dream
- 34. Deep Dream
- 35. Deep Dream
- 36. Artistic Style
- 37. Artistic Style
- 38. Artistic Style
- 39. LipNet
- 40. LipNet
- 41. LipNet
- 45. Let’s Enhance!
- 46. Google RAISR 「Rapid and Accurate Image Super-Resolution」 Upsampling （aliasing artifacts）
- 47. Google RAISR
- 48. Google RAISR
- 49. Google RAISR
- 50. Google Brain's super- resolution technique
- 52. Generative Adversarial Text to Image Synthesis
- 53. Image-to-Image Translation with Conditional Adversarial Nets
- 54. Pix2Pix
- 57. Ratbot and maze
- 58. Ratbot and maze
- 59. Bird-eye camera over the maze
- 60. Digital reward map updating
- 61. Superior learning performance of Ratbots over unenhanced rats
- 62. Embed The Word
- 63. AI can Tell the Difference between Sports
- 64. CaptionBot
- 66. "I think it's a young girl throwing an orange Frisbee in the park," Microsoft's AI will tell you. "I think it's a man jumping through the air doing a trick on a skateboard," Microsoft's AI says.
- 67. AlphaGo 9–15 March 2016
- 68. AlphaGo 9–15 March 2016
- 74. OpenAI Gym
- 75. StarCraft II DeepMind feature layer API
- 77. AI Coopetetion
- 78. Microsoft's AI Tay offends and goes offline
- 79. Open Letter on Artificial Intelligence 08/19/15
- 80. To Benefit People and Society
- 93. Auto-Pilot in China 隧道內部光線微弱，考察車輛在暗環境下的應對策略。 隧道內設置了「白板」障礙物。對於白板的設計，考察 攝像頭和激光雷達的識別能力。
- 100. chaocraig@gmail.com
Editor's Notes
- #73: 1. AlphaGo v13 的網絡層數實際是不夠的。如果按照 AlphaGo v13 的架構，5x5往上面長11層3x3，相當於27x27，看上去夠大了吧？錯，這樣的半徑只有14。 因此，如果大龍的長或寬超出14？ 2. 由於網絡的結構是往上一層層生長，如果隻長幾層，一般不會丟失重要信息，但如果一直長上去，就會越來越容易出現問題。所以，大龍甚至都不用長到14，電腦就已經不一定“知道”自己的大龍是一條聯通的大龍了。 按照AlphaGo v13 的架構，如果大龍只在一端有兩個真眼，另一端就甚至不一定知道自己已經活了（它只會知道自己有兩口氣，而這是網絡輸入告訴它的） 一個誤區：它會傾向於認為氣很多的棋塊就是活的。對於局部死活，這沒有問題，但對於大龍死活，這是不足夠的。 棋塊小的時候，可能出現的形狀不多，而且許多形狀經常在對局中出現，因此容易被神經網絡學會。而大龍越大，其可能出現的形狀就越多，網絡不一定能學會。