Artificial Intelligence

Discuss this lecture:

View comments Leave comment

Alan Turing famously proposed a test of artificial intelligence. What has been achieved? Professor Stephen Hawking has said that real artificial intelligence will mean the end of mankind. Is that a real threat? Are there limits to what a silicon brain might do?

Lecture Date: Tuesday, 13th June 2017 – 6:00pm at The Museum of London, 150 London Wall, London, EC2Y 5HN

Available Downloads

Please feel free to download the lecture slides (powerpoint) and lecture transcripts (Ms Word)

Presentation transcript

Leave a reply

9 thoughts on “Artificial Intelligence

  1. Martyn Thomas CBE FREng says:

    In reply to rhhardin:
    I don’t understand Coleridge’s analogy, but it seems to be arguing for the existence of something other than “matter”. I think Turing dealt with that objection adequately (I quoted and referenced his argument in my lecture).

    Programming is, of course, a results-oriented activity and solving a constrained set of problems is far easier than developing a system with far wider capabilities. So I’m not surprised that your experience is that you stop when you have developed a program that works for the task in hand. I think we all do that!

  2. Martyn Thomas CBE FREng says:

    In reply to Barry Cook: punishment is only one way of teaching that some behaviour is considered undesirable or unacceptable. Machine learning systems do not need to be taught to feel pain, only to be able to distinguish between desired and undesired decisions, which is what the training data does.

  3. Barry Cook says:

    One highly desirable aspect of real intelligence that I have not seen discussed is that people (and animals) moderate their behaviour by understanding the concept of punishment for wrongdoing. All our guidelines (laws) are framed in terms of undesirable sanctions for transgressions. Is there any way to similarly control AI? Does AI recognise outside judgement? How can an AI system be punished? So far, I suggest, we have decision-making/reasoning machines but they cannot be called intelligent until they understand the concept of wrongdoing (harm to others) and punishment as it applies to themselves – and behave accordingly.

  4. rhhardin says:

    Coleridge in Biographia Literaria, chapters 5-8, covers the impossibility of machine intelligence with an argument that still works, roughly in summary that matter has no inwards. You remove one surface only to meet with another. It’s the wrong kind of thing to be intelligent.

    As a programmer, I’ve found, you start out trying to do something finally intelligent and wind up solving the problem by finding something that works on the specific problem that isn’t intelligent at all.

  5. Raymond Kershaw says:

    Thanks. I confess I do not know chess. If you wish I can find explanations of how backgammon neural nets are developed (the original one was at IBM about 20 years ago) and send them to your personal e-mail (if you supply it). (You are correct that optimal doubling is as important as optimal checker play.)

    I wonder whether the random element makes machine learning as applied to backgammon more akin to some of the practical applications which you discussed in your lecture than machine learning as applied to chess or Go?

  6. Martyn Thomas CBE FREng says:

    I think you underestimate the complexity of chess! Each player has 16 pieces, of 6 different types that can make different sorts of moves, and a single move of a single piece may result in 8 different board configurations. But I agree that backgammon has a lot of complexity beyond the 21 dice throws and that decisions about doubling are (in my experience) at least as important as decisions about moves, over the course of a multi-game match.

    Thanks for the information about eXtreme Gammon. I look forward to investigating it further.

  7. Raymond Kershaw says:

    Brute force was tried with feeble results, except in no contact end game positions. Unlike chess, there is not just one move to consider on a player’s turn but 21. The multiplication gets huge. And there is also the doubling option to consider. The program trained by playing itself a large number of times and gradually adjusting the weights given to different patterns of the checkers, as a feedback from the outcome (win or lose).

  8. Raymond Kershaw says:

    In your lecture on Artificial Intelligence which I attended today, you reference the application of machine learning (neural nets) to chess and Go. Another successful application is to backgammon where the challenge is to look ahead up to four moves, allowing for the 21 possible rolls of two dice (15 non-doubles and 6 doubles) for each move. The strongest program, eXtreme Gammon, is better than any human player.

Leave a Reply

Your email address will not be published.