|
Post by Hannes Vilhjalmsson on Jan 27, 2011 5:48:38 GMT -5
The third paper we'll discuss will address the question of what artificial intelligence is in a very practical manner. It describes the field of AI with several examples, some of which you have seen in the class so far (e.g. search) and some which you will see later in the class (e.g. learning). One interesting thing about this paper is that it is written by a philosopher for those who are studying the brain, e.g. not necessarily for computer scientists. A nice thing about that is that it gives a very concise picture of a field within computer science with relatively "fresh eyes" and is also able to use concepts such as mind and consciousness knowledgeably. The paper is from the chapter "Artificial Intelligence" in a book called "Matter and Consciousness" by Paul M. Churchland. It is a revised edition from 1993. There have been advances in AI methods since then, which we will cover in the course, but the fundamentals are the same. You can retrieve the paper here: programming_intelligence.pdf
|
|
|
Post by carmine on Jan 28, 2011 10:05:42 GMT -5
- A brute-force algorithm should create a perfect player... In some games we need to create several levels of difficulties for a player. How is it possible manage this problem? Do we need to apply a heuristic model or just make some changes on the brute-force algorithm? What is the better approach??
- How much can technologies like CUDA help us in problem with an high time consuming? I know that it is very useful in problems about recognizing something.. But Is it really useful in problems that requires a huge knowledge like NLP?
|
|
|
Post by kristjan on Jan 30, 2011 3:08:23 GMT -5
About the problem of facial recognition on a photo i was wondering if it would be a good approach to recognize in which direction the face is facing and have a computer algorithm reconstruct the face so it is looking straight forward before trying to find out if it is a face that it recognizes to make the process of recognizing easier.
|
|
|
Post by helgil08 on Jan 30, 2011 7:51:47 GMT -5
-if vision is one part of ai, natural language manipulation one, purposive behavior and problem solving one, learning one and self conciousness one, are there any other parts in ai the author didn't mention.
-are there any well known methods to create heuristics or is it an ad hoc problem or a general problem solving skill that's used each time. is there any science behind it or what kind of mathematics are used.
|
|
|
Post by finnur on Jan 30, 2011 10:23:49 GMT -5
-in the end of the learning part, he talks about how new approaches to representation and manipulation of large amounts of information have recently procured very striking "learning procedures" but doesn't say anything more about it, what are those?
-instead of programming a computer to understand natural language manipulation, wouldn't it be better to try and give the computer the capability to learn it by watching others talk, and learn language manipulation like that?
|
|
|
Post by sigurdurjokull on Jan 30, 2011 11:18:02 GMT -5
He was talking about the strategy of taking on individual parts of the brain and trying to program AI versions them and the disadvantage of it since for instance something like vision depends largely on the knowledge an information processing system has. I was thinking about all the things the human brain can do and how important interconnection must be in everything the brain does. For instance the problem of facial recognition probably becomes a whole lot easier once you are able to separate the person from the background, which is something AI programmers probably do. What AI programmers probably don't have, but the human brain probably does have, is just a lot more of these kinds of tricks, our visual processing is a huge collaboration of smaller brain areas which work together, and the visual processing has connections from other parts of the brain. Helping the visual processing know what to expect, and making the process more precise. And since the design process is evolution, the strategy is just try every connection and throw out the bad ones, and we have to think of all of these connections which is a lot harder. With all kinds of tricks like that, i can see why the brain has such an advantage.
It seems that integration and interconnection of information is really important, is this something that is being researched? Is there hope of learning something like the fundamental rules of how the brain operates and forms connections from neuroscience or from genetics?
|
|
|
Post by Þorgeir Karlsson on Jan 30, 2011 11:45:37 GMT -5
This paper was written in 1993 so what advances have been made in connectionism / PDP? Could further advances in this field mean that in the future, computer architecture will be unrecognizable to us?
Is memory a problem in machine learning? i.e. if a chess computer or an intelligent agent is saving more and more data to "learn" from would it not become slower and slower in its evaluations as the percept sequence becomes larger?
|
|
|
Post by Ásgeir Jónasson on Jan 30, 2011 12:17:48 GMT -5
1.
The article mentions Tic-Tac-Toe as an example of an easy game to play using brute force, and chess as an extremely difficult (or impossible) one. What are some of most complex problems computers can solve using brute force (in a reasonable amount of time and space) ? Do you think we might revisit brute force techniques again as hardware improves, for example if quantum computers live up to their expectations ?
2.
"If machines do come to simulate all of our internal cognitive activies, to the last computational detail, to deny them the status of genuine persons would be nothing but a new form of racism"
I know I'm venturing into Sci-Fi area here, comes with the territory I guess, but what do you think an AI machine's status in our world would be, for example, regarding criminal law ? We seem to classify our violent crimes and punishment in regard to the intelligence of the victim and the offender, if they are of separate species. And it works in both ways. A dog bites a man and is killed for it, and a man killes a dog and isn't punished as much as if he had killed another human being. Where do you think machines would come in this ranking? If they become more intelligent and self aware, would it be a greater crime to harm one of them, or would it not be punishable at all ?
|
|
|
Post by lorenzo on Jan 30, 2011 13:23:42 GMT -5
- I think that the memory space for the machine learning is not a problem becouse the machine is going to modify the rule that it has on the basis of the data and the result of the current rule so in that case is not a problem. Instead for the performance of image recognition that is a big problem becouse you have to check if you all the data for know you can recognize the object.
- In my opinion i think that a machine can't reach the same performance of us in image recognition or in the other areas mentioned by the author.
|
|
|
Post by Helgi Siemsen Sigurðarson on Jan 30, 2011 15:13:06 GMT -5
Could cloud computing get us closer to making brute force possible ?
Could combination of many type of sensors such as sonar and heat vision make AI capable of seeing better then us or are they still limited by the hardware ?
|
|
|
Post by thorsteinnth on Jan 30, 2011 15:27:27 GMT -5
Í sambandið við að vélar geti "lært", það er sagt að það taki upplýsingar, og geymi það í minni til að geta sótt það hratt aftur, er það ekki bara.... minni? Ekki lærdómur?
Ef tölvur myndu í raun þróa meðvitund, hvernig gætum við vitað að þær væru í raun með meðvitund en ekki bara háþróaður hermir?
|
|
|
Post by hordurh10 on Jan 30, 2011 15:48:57 GMT -5
An experienced chess player can in many cases judge a position on a chessboard as won, lost or draw in a matter of seconds (is: stöðumat). The method is closer to learning than searching for a goal. If we assume that all positions of all played chess games were available, could we write a decent chess program almost entirely based on the data?
In light of what Churchland says about character recognition "Plainly, this system is inflexible and can easily be victimized." is it justifiable to use the technique in critical cases such as counting votes in a national election?
|
|
una
New Member
Posts: 12
|
Post by una on Jan 30, 2011 16:07:17 GMT -5
Í sambandið við að vélar geti "lært", það er sagt að það taki upplýsingar, og geymi það í minni til að geta sótt það hratt aftur, er það ekki bara.... minni? Ekki lærdómur? But isn't it also possible to say that that's what learning really is ? Remembering an answer to some kind of problem, or a reaction to some kind of situation. For example, if we touch a hot stove, we hurt ourselves, and that goes into our memory, so next time we're faced with a situation like that, we remember that it hurt, so we don't do it. Or in games, if we die at a certain place, we'll remember it, therefore learn to not make the same mistakes... --- Could it be said that artificial intelligence ( in terms of natural language manipulation and socialising in general ) is currently closer to autistic individuals than "normal" people? In the sense that they are socially impaired and the world can be hard to comprehend, while in some cases they possess a hightened ability to memorize but harder time applying those memories to use. Just a thought...
|
|
|
Post by Hrafn J. Geirsson on Jan 30, 2011 16:32:38 GMT -5
1. "Computers are very good at number crunching, theorem proving and list searching, while they are bad at facial recognition, scene apprehension, sensorimotor coordination, and learning."
This is not surprising since computers were originally built to perform these functions. If we are to make a machine, capable of exhibiting sophisticated animal type intelligence, is there a possibility that we need to unlearn some of the mathematical ways of computing to make a turn in the right direction ?
2. The author states that it is naive to believe, that consciousness is a single, uniform phenomenon to be captured. Does that mean that there are many different types, maybe even degrees of consciousness ?
|
|
|
Post by Eiríkur Fannar Torfason on Jan 30, 2011 17:04:52 GMT -5
- A brute-force algorithm should create a perfect player... In some games we need to create several levels of difficulties for a player. How is it possible manage this problem? Do we need to apply a heuristic model or just make some changes on the brute-force algorithm? What is the better approach?? I can think of two ways of implementing a turn based game playing agent with varying skill levels. First, you could vary the search depth so that the agent only considers X rounds into the future. You still need some sort of an evaluation function to assign a score to each branch so that the "best" move can be selected but the search could still be brute-force (explore all branches up to the depth of X). The other way is to have different evaluation functions depending on the difficulty level. The more naive/simplistic the evaluation function, the easier it becomes to beat the agent. You can test your reversi skills against different kinds of agents (bots) on this page here: www.site-constructor.com/othello/
|
|