csci 21 215 soc ocial eth thical iss issues in in com
play

CSCI 21 215 Soc ocial & Eth thical Iss Issues In In Com - PowerPoint PPT Presentation

CSCI 21 215 Soc ocial & Eth thical Iss Issues In In Com omputing Class 22 (some) notes on AI Not otes Guest Speaker Tomorrow Guest Speakers Tuesday Researchers at our very own university who have worked with social media


  1. CSCI 21 215 Soc ocial & Eth thical Iss Issues In In Com omputing Class 22 – (some) notes on AI

  2. Not otes • Guest Speaker Tomorrow • Guest Speakers Tuesday • Researchers at our very own university who have worked with social media and privacy concerns • Homework 4 posted • You can you any of the links on today’s slides for homework 4

  3. Revie iew • How many software freedoms are there? • What’s the difference between GPL and Copyleft? • Also, James was correct, BSD License is the license concerning the use of the creator’s name.

  4. Net t Neu eutrali lity ty Ru Ruli ling – 6/1 /14/2016 • Broadband is now classified as “common carriers”, meaning FCC can enforce laws to keep it open and neutral • Service Providers (including AT&T) are vowing to appeal to the supreme court

  5. Tur urin ing Tes est t - Orig rigin inal • Interrogator has just 5 minutes to chat with human/chatbot • AI System passes if judges get fooled >30% of the time • Chatbot Eugene Goodman made news in 2013 for passing Turing contest at University of Reading

  6. Tur urin ing Tes est t - Upd pdated • 1990 Hugh Loebner updated the Turing Test • Conversations should last 25 minutes • Should be 4 different judges • AI System passes if judges get fooled >50% of the time

  7. Tur urin ing Tes est t - Th The e Loe oebner Pri rize • Contest consists of 4 rounds • Each round, the 4 judges interact with two entities using a computer terminal, one human and one AI. • After 25 minutes of questioning the judge must decide entity is the human and which is the AI. • Gold Medal, if a system can fool judges on more than just text level • Silver Medal if can fool more than half judges • If no winners, Bronze medal awarded to AI that fools most judges http://aisb.org.uk/events/loebner-prize

  8. Arti rtific icia ial In Intelli ligence e – Lea earnin ing & Som ome His istory • 1964 ELIZA - Created by Joseph Weizenbaum of MIT. • Original Chatbot that picked up on key words and imitated a therapist . • For Example: ELIZA would respond to a question such as “My mother hates me” with “ Who else in your family hates you ?” • Weizenbaum went from a strong proponent of anthropomorphized AI to a critic after test subjects started to open up emotionally to ELIZA, as if she were a real person.

  9. Arti rtific icia ial In Intelli ligence e – Lea earnin ing & Som ome His istory 1957 Perceptron - using a large number of very simple linear classifiers creates a strong model

  10. Progress in machine classification of images • Red line - the error rate of a trained human Microsoft Research- outperforming humans

  11. Arti rtific icia ial In Intelli ligence e – Lea earnin ing & Som ome His istory • 1990s – Statistical AI and Big Data – Recognize patterns • 1996 – Deep Blue – won chess match against world champion Kasparov • Won first of 6 games, lost 2, and tied 3 • 1997 Rematch, Deep Blue won 2, lost 1, and tied 3

  12. Arti rtific icia ial In Intelli ligence e – Lea earnin ing & Som ome His istory • 2011 – Watson and Jeopardy – understands natural language (enough to beat the top 2 jeopardy champions while not connected to the internet ) • 2013 – Watson goes commercial – Utilization Management Decisions in lung cancer treatment • http://www.ibm.com/smarterplanet/us/en/ibmwatson/offerings.h tml

  13. Goo oogle e Dee eepMin ind • AlphaGo v Lee Sedol at GO competition • AlphaGo wins 4 -1

  14. Goo oogle e Dee eepMin ind • British Artificial Intelligence company, acquired by Google • Being used in UK for kidney disease treatment • http://www.theguardian.com/commentisfree/2016/may/04/googles- deepmind-shouldnt-be-sucking-up-our-nhs-records-in- secret?CMP=fb_gu • As a side note, on medical data security • http://money.cnn.com/2014/08/20/technology/security/hospitals- data/

  15. Type pes of of AI 1. Artificial Narrow Intelligence (ANI) • REALLY good at one thing (i.e. playing chess) 2. Artificial General Intelligence (AGI): • Intelligent on par with human • Could pass Turing test 3. Artificial Super Intelligence (ASI): • An agent who possesses intelligence far beyond even the most gifted human. • http://www.nickbostrom.com/superintelligence.html

  16. Disc iscussio ion • How can we balance the good of giving all data to a system like Watson and Google DeepMind with privacy concerns? Should we?

Recommend


More recommend