Transcript Document
Artificial Intelligence Natural Progression • We can solve logical problems • We can build machines to solve general problems • Can we make the machine replace us? • Roots in science fiction – Androids • An android is an anthropomorphic robot - i.e. a robot that looks like a human. For example Valerie • Turing test – Can a computer behave in such a way that someone interacting with it remotely cannot differentiate between it and a human being? What do humans do well • Process information – Language – Images • • • • • Reason Create Relate to other humans Make intuitive decisions Experience and exhibit emotion – Enjoy entertainment What do computers do well • Process data quickly • Handle large amounts of data Computers simulating humans • Have to understand the process • Build a good model for the process • Resolve complexities Example: Chess • Number of possible unique chess games is 10120. • In 1957, artificial intelligence pioneers Herbert Simon and Allen Newell predicted that a computer would beat a human at chess within 10 years. • BELLE, a chess program by Ken Thompson and Joe Condon, became the first computer to be awarded the title of US chess master, in 1983. • BELLE didn’t try to do what a human would do. Instead, BELLE took advantage of what computers do well. • In May 1997, IBM's Deep Blue Supercomputer played a fascinating match with the reigning World Chess Champion, Garry Kasparov and won 3 ½ to 2 ½ Example: Chess • Does this count as AI? – Computer beating best human – Computer not playing as human would Other Successes • Language understanding – Eliza 1967 • Speech recognition – Dragon Naturally speaking – Phone company systems – Voice mail systems • Language translation Difficulties • Humans are complex beings and understand – Language issues • Time flies like an arrow • Fruit flies like a banana – Searching for tables • Table • table chair • 98% correct is often not enough – 40 typos in a page • The search space can get very large Techniques • Build a search tree to model the state space • Find good methods of evaluating possibilities • Use your evaluation methods to prune the tree First steps in TicTacToe First move center corner side center . . . . O. . . X . . . . X. . . O . . . . XO . . . opp. corener corner side adj. corner adj side opp side Turing Test • A person tries to distinguish between a man and a woman (responses over a typewriter). If you replace one by a machine, does the game change? • A person and a machine are behind a curtain. An interrogator sends questions to each, and receives text message answers from each. Can the interrogator tell which is which? • Turing predicted machines would be able to be programmed to pass the Turing (or imitation) test • 50 years of philosophical debate have followed Strong AI vs. Weak AI • Strong artificial intelligence states that a computer with the right program can in fact be mental (like a human being) • Weak artificial intelligence just aims to solve problems, not necessarily to be mental or model human behavior. Weak AI • A lot of it was based on rules – Rule-based expert systems: • If (Smoker) and (White Male) then (Chance of Lung Cancer by Age 60 > x) – Rule-based categorization • If news story mentions baseball or hockey or … at least three times, then it should be categorized under Sports • Recently, greater focus on probabilistic and statistical methods Probabilistic Inference • Xi are diseases, boxes are symptoms • Prob. of each disease given each symptom is known • Given new set of symptoms, infer probs. of diseases Machine Learning (ML) • An increasingly popular subfield of Artificial Intelligence • Usually aimed at performing some particular task, not building sentient machines • Design of systems that improve (hopefully) as they acquire knowledge or experience ML slides taken (and modified) from Ken Williams’ tutorial Typical Machine Learning Tasks • Clustering – Grouping similar items together • Categorization – Of text, images, news, into categories • Recognition – Speech, Voice, Handwriting • Game playing – Chess • Autonomous performance – Robots that play soccer or navigate themselves Typical ML Tasks • Clustering Typical ML Tasks • Categorization Typical ML Tasks • Recognition Vincent Van Gogh Michael Stipe Mohammed Ali Ken Williams Burl Ives Winston Churchill Grover Cleveland Typical ML Tasks • Recognition Little red corvette The kids are all right The rain in Spain Bort bort bort Typical ML Tasks • Game playing Typical ML Tasks • Autonomous performance What Do ML Systems Do: News Categorization • Start with a set of news articles, that have been manually tagged with categories (Sports, Politics, Business, Baseball, Customer Loss) • Build a statistical model to represent the characteristics of each category, based on the above “training data” • Upon a new article (“test data”), test it against the models for the categories to see where it fits best • What is learned is the model from the training set. Then the learned model is applied to the test set. • E.g. Voice recognition When is ML useful? • When you have lots of training data • When you can’t hire enough people, or when people are too slow • When you can afford to be wrong sometimes – ML systems have accuracies ranging from 50 to 85% • When you need to find patterns What are the alternatives? Back to AI and the Human Brain: Work in other sciences • Neural Nets – Build models of machines that think and learn • Brain mapping – Determine what clumps of neurons do – Eventually map individual neurons Questions Worth Exploring • State of art – State of the art in AI – Technology in general • Nature of thinking – How does the brain do it? – Must a machine duplicate the brain? • What is learning? • Philosophical – What makes a machine human • Science Fiction State of the Art What would happen if we kept asking the robot why it just gave the answer that it gave an infinite number of times. Could it explain? For example: Input: How are you today? Output: Good. Input: Why "good"? Output: I'm having a lovely day. Input: Why? Output: I don't know. Input: Why don't you know? ...and so on State of the art • Is there really any way that a computer can create random numbers or is it always just numbers that appear random to a human being? Nature of Thinking Do machines really think or just follow orders? • A machine that does AI has a set of rules and guidelines from which it may not deviate. Do these rules make the machine nothing like the brain, seeing as the brain has no rules or guidelines to work within, or can these rules and guideline be equated to the capacity an individual's brain has to learn and make sense of problems, ideas, objects, ect.? Can machines learn to think? • Are modern definitions of a computer "thinking by itself" still dependent on continual human programming as the only means of progress, or are there computers somewhere in the distant or even near future that can actually be programmed to program themselves based on input of users or even just experiences (the way that human beings learn and are taught to think)? Nature of thinking (cont) • Can machine language and thought be superior to human thought? If so, what does this say for the future of society? • If scientists can ever figure out the exact code for the brain, would it be possible to simulate it in a strong AI computer? Learning • Is it possible for a computer to learn? Can it adapt to unfamiliar situations and "reprogram" itself? • Because a machine cannot have experiences, how can a machine learn from experiences and acquire the same type of conditioning that humans acquire throughout their lives? • Could a machine ever write a sonnet with the same intensity and meaning as a human poet or truly enjoy the taste of summer strawberries and cream? • Can machines ever be programmed to experience emotions? Philosophical • We now accept that a computer can beat a chess champion, but we don't easily accept that a computer could experience emotions in the same way that we do. Is this resistance a general belief that human consciousness is so unique that it could not be replicated by humans? Or is it another idea that will be commonly accepted in a few more decades? • If a computer can map out responses through a network of possible answers and connecting ideas, would that make it more human? • See articles on Turing’s theories and their its implications/rebuttals: http://www.turing.org.uk/turing/scrapbook/test.html, http://cogsci.ucsd.edu/~asaygin/tt/ttest.html Computers vs. Humans • If we create a machine that is a replica of a human brain, using electrons in place of neurotransmitters, is that machine a thinking machine? Even if it cannot think on its own and needs to be able to read the brain activity of a person in order to function. If it is able to follow human neurotransmitter patterns is it thinking or merely copying? • Is programming a "life history" into a computer possible? • Is if possible for a computer to simulate creativity? Let’s define creativity as the ability to figure out a solution without the inputs necessary to draw that conclusion from pure logic. Science Fiction • Do you think that it will be possible one day to store information in the human's brain or to educate humans by using computer programs? (Like in the movie “The Matrix,” when the main character gets fighting techniques and jump programs stored in his brain) • How far are we from robots that look and move in ways very similar to humans? • Do you think that if in a point in time we develop machines that could think and be human-like then they would represent a danger for humans?