1. What is Artificial Intelligence? Introduction to Artificial Intelligence Week 2, Semester 1 Jim Smith

2. Overview of this week’s topics. What is Artificial Intelligence? The Turing Test. Searle’s Chinese Room Argument. Strong vs. Weak Artificial Intelligence.

3. What is Artificial Intelligence? "The exciting new effort to make computers think" - Haugeland,1985. "The automation of activities that we associate with human thinking, activities such as decision-making, problem solving, and learning" - Bellman, 1978. "The art of creating machines that perform functions that require intelligence when performed by people" - Kurzweil, 1990. "The study of how to make computers do things which, at the moment, people are better" - Rich and Knight, 1991. "The study of mental faculties through the use of computational models" - Charniak and McDermott, 1985. "The study of computations that make it possible to perceive, reason, and act" - Winston, 1992. "Computational intelligence is the study of the design of intelligent agents" - Poole, 1998. "AI is concerned with intelligent behaviour in artefacts" - Nilsson, 1998.

4. So does intelligence require thinking? Definitions focus on how to make: – systems that reason or act like humans, and – systems that think and act rationally. Raises the big question: what is intelligence? I would add a second question: do we need human intelligence?

5. The Turing Test An empirical approach designed to provide a definition of intelligence, based on the measure: "Does a system act as if it is human?“ Proposed by Alan Turing in 1950. – He thought that computers would pass it by 2000. – Annual competition with big money prize. – still not passed.

6. Turing Test, version 1. Stage 1 Interrogator sits in a room with a keyboard and screen. Man and woman sit in separate rooms. Task is to ask questions to work out which is female. Woman co-operates but man tries to fool interrogator. Stage 2. “Intelligent” computer takes the place of the man. Interrogator has to tell which is which.

7. Turing Test, version2 Simpler version where the interrogator just has to distinguish AI system from a human. Turing says a machine has passed the test if: “an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning.”

8. So how do you trick a computer? Think about this in the tutorials: make a list of questions to ask and then try it out at the jabberwacky site. http://www.jabberwacky.com/ Jabberwacky is written by the guy who has won the last two Loebner prizes. This is an annual competition to find the best attempt to pass the Turing test.

9. What does the Turing Test not do? Turing, A.M. (1950). "Computing machinery and intelligence." Mind, 59, 433-460 -- Quote from Section 6. The original question, 'Can machines think?' I believe to be too meaningless to deserve discussion.

10. An Aside: how to reference. Early on we included a load of quotes as “blah blah blah,” Smith 2007. this is a good start, because it means that we are attributing other people ideas to them, – not passing them off as our own (plagiarism). – our necessarily agreeing with them. but what if someone wanted to see the quote in context? – not everything is google-able.

11. how to reference 2: quoting. A couple of slides ago I gave another quote from Turing. This was better style because: I used text style to show that I was quoting directly. and I provided a full reference that you could go to the library and find or order it.

12. How to reference 3: Summarising. If I wanted to say summarise an argument I might say something like: Turing believes that his test is a practical one, and that to use phrases like “thinking” risks descending into endless debates about precise meanings of words [1]. then at the end of my document I show what this reference is so that the reader can go and find it if they want to read more, e.g.: 1. Turing, A.M. (1950). "Computing machinery and intelligence." Mind, 59, 433-460 -- Quote from Section 6.

13. Materials and referencing. you will receive more detailed instructions elsewhere about academic conventions concerning, referencing plagiarism etc. This year I will mostly give you papers to read and on-line materials that provide overviews or summaries – see how they reference other people. In later years you will be expected to read more original texts yourselves.

14. Anyway, back to the topic…

15. Passing the Turing Test. A machine would require at least the following: – Natural language processing - to communicate in natural language, such as English. – Knowledge representation capabilities - to store what it knows. – Automated reasoning - to use this stored knowledge to answer questions and draw conclusions. – The ability to learn - machine learning enables systems to adapt to new situations and detect patterns in phenomena. – and possibly computer vision - to perceive the world. – robotic capability - to explore and/or manipulate the world.

16. Criticisms: Searle’s Chinese Room. In 1980 John Searle start a huge and fierce debate about the nature of intelligence and whether Turing’s test was an adequate test. His argument is based around the idea that a computer could pass the test if it: – recognised the questions (syntax), – without needing to understand them (semantics). – Searle, John. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3 (3): 417-457.

17. The Chinese Room: 1. A number of people approach a room and push slips of paper through a hole in the wall. After a while each slip comes back out. The people look at them and seem satisfied. It turns out that the original slips have questions written on them in Chinese. They return with responses in Chinese. Is the room/system showing intelligence?

18. The Chinese Room 2. Searle suggests that inside the room is non- chinese speaking person and a lot of books. The person looks at the slips and then goes to a book which tells him/her what to write. – e.g. “if you see XYZ then write DEF”. His argument is that the person is just recognising the symbols (syntax) but not understanding what they mean (semantics).

19. Chinese Room 3. Searle’s argument is basically that the Chinese room, and by extension other AI systems: – appears to solve tasks, – but are really just following complicated programmes, – without understanding what they are doing. He would argue that in order to create true AI (Strong AI) a machine needs to experience the world more directly. Many have argued this needs a kind of consciousness.

20. Strong vs. Weak AI. Strong AI is the attempt to create machines that have human like intelligence, e.g.: – the ability to possess a model of reality. – reason on the basis of that model to make predictions or plan actions. Weak AI is the attempt to create machines that perform tasks that would seem to require human intelligence.

21. Strong AI. Searle and many others have engaged in fierce debates about whether “Strong” AI is achievable. – still some research going on, see wikipedia. Most approaches assume that the “world model”, and reasoning use logic based on rules and facts. This is the Symbolic Processing approach.

22. Weak AI. Meanwhile many others have just got on with the problem of making useful machines using a weak AI approach. – Could use symbolic reasoning restricted to a very limited subset of a pre-programmed world. Or approaches with no world model at all: – Often based on biological or physical metaphors. – subsymbolic processing / Computational Intelligence.

23. Strong vs. Weak methods. A separate distinction can be made according to how much knowledge the systems are given: Strong methods rely on using knowledge about world and specific tasks. Weak methods don’t rely on knowledge or understanding of the world or the problem at hand.

24. Summary. The question “what is AI” can force us to think a lot about what it is that makes us intelligent. Are other animals intelligent? Does intelligence = general problem solving? BUT just because we use our intelligence to do lots of useful things, it doesn’t mean that computers need do so.