Slashdot Mirror
Stanford 'Intro To AI' Course Offered Free Online
An anonymous reader writes "IEEE Spectrum reports that Stanford's CS221 course 'Introduction to Artificial Intelligence' will be offered online for free. Anyone can sign up and take the course, along with several hundred Stanford undergrads. The instructors are Sebastian Thrun, known for his self-driving cars, and Peter Norvig, director of research at Google. Online students will actually have to do all the same work as the Stanford students. There will be at least 10 hours per week of studying, along with weekly graded homework assignments and midterm and final exams. The instructors, who will be available to answer questions, will issue a certificate for those who complete the course, along with a final grade that can be compared to the grades of the Stanford students. The course, which will last 10 weeks, starts on October 2nd, and online enrollment is now open."
When asked how they would deal with ten thousand students, Professor Thrun replied:
"We will use something akin to Google Moderator to make sure Peter and I answer the most pressing questions. Our hypothesis is that even in a class of 10,000, there will only be a fixed number of really interesting questions (like 15 per week). There exist tools to find them."
Just like personal robots, flying cars, and apartments on the moon, a worth while dream.
Proposed:
A software program which can successfully pass this course.
Related: Turing Test
My first Journal Entry ever, in 8 years! http://slashdot.org/journal/365947/aphelion-scifi-fantasy-horror-poetry-webzine
The actual website for the course says "The class runs from Sept 26 through Dec 16, 2011." http://www.ai-class.com/
Presumably since it's online, anyone can take it. The article says only Stanford students will get credit, but if you can convince your own university to give you credit for it ... do it!
I dont buy that. There were times in grad school when a class of 20 students generated enough questions on a topic that threw the instructors schedule out of whack. I know this is not grad school but I am assuming there are enough good students in Stanford itself and most people who will sign up voluntarily will be the ones who are interested about it. I still love the idea though. Although, it makes me wonder how the students feel about it. Stanford is pretty expensive. They have paid all that money only for coming to class now, given that the exact same class material and the instructors as well are available to anyone for free ?
about AI is a little too recursive for me...
"When asked how they would deal with ten thousand students, Professor Thrun replied, 'We'll let Skynet handle the sorting and choose the best questions'"
It doesn't look like it's a credit course for anyone who isn't a registered Stanford student. They give you a certificate of completion (Which, when combined with $1.50 will get you a cup of coffee), but not actual course credit. On the other hand, this is a course taught by two of the top researchers in the field. It's probably worth it just to learn something. I'm seriously considering this. I don't know a lot about coding AI, beyond some really high level theory; and while I'm sure that a ten week course with 10,000 of my closest friends won't make me an expert... It could be fun.
I don't need a million points of light, just two points of multi-mode fiber and a 10 Gig-E router.
Now I don't have to live in Massachusetts to learn me about some artificial intelligence!
the preceding comment is my own and in no way reflects the opinion of the Joint Chiefs of Staff
http://ocw.mit.edu/index.htm
http://www.wikipedia.org/
http://www.khanacademy.org/
http://www.ted.com/
Standford isn't first to this game, but I still applaud them.
http://blindscribblings.com - Tasty pop-culture in conceptual fashion.
I already learned (thanks Wikipedia!) that Stanford is not part of the ivy league, though :)
Can I earn college credit that will transfer to other institutions? That's what I want to know.
James Tiberius Kirk: "Spock, the women on your planet are logical. No other planet in the galaxy can make that claim."
I don't think that e-book means what they think it does :)
CS221 is the introductory course into the field of Artificial Intelligence at Stanford University. It covers basic elements of AI, such as knowledge representation, inference, machine learning, planning and game playing, information retrieval, and computer vision and robotics. CS221 is a broad course aimed to teach students the very basics of modern AI. It is prerequisite to many other, more specialized AI classes at Stanford University.
sounds like electronic human brains is the goal to me
The university is more than just a course lecture, its an environment. If I didnt know something I just asked across the coffee table or a door down the hall.
Haven't read his AI book "Artificial Intelligence, A Modern Approach".
But about 20 years ago when I was really into Common Lisp, I read his book "Paradigms of Artificial Intelligence Programming: Case Studies in Common Lisp". It was one of the best books I had ever read. Lots of fantastic examples and code.
Makes me think I should get his "modern approach" book. Maybe think about the online course.
I'll see your senator, and I'll raise you two judges.
Interesting. Now is a good chance to find out for one-self if the famous USA universities really are as elite as their reputation. And I will learn basics of AI at the side! I already learned (thanks Wikipedia!) that Stanford is not part of the ivy league, though :)
Actually you've been able to do this for a while. Check out iTunes U, http://www.apple.com/education/itunes-u/what-is.html.
:-)
FWIW, Ivy League is a marketing gimmick.
It is about time the universities go the way of the book, music, news, encyclopedia and information industries in which the Internet has brought down prices significantly. There is no justification to the huge amounts of money the universities charge for the education they provide and for the learning materials. This course is a live demonstration of how it can be done for pennies. The only thing that should cost more then a few dollars is final testing of a course which can be done for about $50.00 per course. Beyond that if remote testing is used it will be very hard for any one to get a whole degree by cheating on the total amount of courses needed to graduate. A few random tests on key subjects where you have to be present physically are more then enough to put an end to any shenanigans. The only reason it has not happened yet is psychological, people (those who study and those who hire) being conservative by nature want a degree from "a well known establishment". (Yes some courses need labs and cadavers but they are a small minority)
Three words: Graduate Teaching Assistants.
If the content of this class is exactly the same as the "real" version, and at the end you are evaluated on the grading curve right alongside "real" students... then you have to question why the cost of "really" being a Stanford student is $55,385 per year, while the cost of receiving the same product without the formal diploma is $0.
How much of the expense of modern university education today is actually tied to the core product, and how much is simple sociology? That is, only a certain percentage of society can be in the "elite" ranks by definition... and so elite institutions must price themselves accordingly to maintain the appropriate exclusion.
Don't expect to be doing a lot of coding AI in an intro course. If they're anything like my AI classes, they are 9 parts conceptual, 1 part pseudo-code. Maybe you'll get to write a bit of Lisp in the meantime (AI people think Lisp is the language of the gods, and trying to change that is like prying CoBOL from bankers or FORTRAN from Mech-E people).
The field of AI is no longer focused on creating humans brains as far as I've learned from my studies. They did dream big back then when the field first came to be, but the complexity of the problem became apparent. It's simply, currently, not possible.
There is planning, search and logic AI, which finds the best possible plans for different problems, and is often used in manufacturing. Such as designing computer chips, or for instructions to robots or cranes that builds, sorts or package. AI is capable of approximating solutions to problems that cannot be done through algorithmic means; as such AI often deal with problems in NP.
Another field is game AI, which I know most about. There's a plethora of sub-fields here. The traditional game AI dealt with solving games, and has influenced many games such as chess. (AI hasn't solved chess, but found many end games that humans did not know, and found solutions to end games that humans have theorized about for over a hundred years) Modern game AI concerns itself with AI for video games. The goals are many. Fun and challenging opponents. Autonomous opponents that learn during play and gain new knowledge. Procedural content generation in respect to the player and much more. Not that much has been done in the industry, but in the field there's a lot of focus on machine learning techniques that learn the games themselves based on some criteria set by the creators.
I haven't read anything about AI that attempts to be human-like in the sense they pursued earlier lately. I've read several times however, that the Turing test is faulty and should be ignored; it serves no purpose in the field. The new purpose is to create machines that can do some task, and do it well. If its deemed intelligent by humans is of no consequence. If it does a job better than a human, then it is an advance. That it is worse than optimal is a strength, because as I said, the problems often dealt with are not solvable optimally. (At least not until quantum computing, albeit I know nothing about how that works; it seems to be another new dream, so if its like the dream of AI in the beginning, it will probably not solve all, but just make advances)
01 REDEFINE REALITY.
TFS says that the online enrollment is open. I couldn't find any way to enroll, only a page where you can enter your name and email to "sign up [...] to receive more information about the online version when it becomes available". Am I missing something? Does anyone have a link to where you can truly enroll for the free version of the course?
A successful API design takes a mixture of software design and pedagogy.
I strongly believe it is possible - it just takes actual funding!
I have cynically remarked that it is a "racial fear" which prevents us from really funding the research it would take to really hit the singularity. That, and our current greed and lust for power playing wargames.
We could have had it by 2012. Taking the famous marker of 9-11, instead of the disastrous failed decade, if we had poured that Trillion into a broad research pyramid, we'd be there. 75 teams of 25 people working for 10 years - yep. Piece of cake. But no, we had more fun invading the wrong countries and groping fliers.
My first Journal Entry ever, in 8 years! http://slashdot.org/journal/365947/aphelion-scifi-fantasy-horror-poetry-webzine
I took an A.I. course in college my senior year and one of the very first assignments we were tackled with was the classical N-Queens problem and different variants of the problem was introduced. Since then I have been intrigued by A.I., specifically in terms of games like starcraft so this will be very helpful.
One thing that i havent looked at closely was that it was noted on one of the links that there were some prerequisites that the students had to meet before taking this course. I'm not sure what those are exactly but i will take a look.
The text for this book is over $100... I know textbooks are expensive, but for a course that will not lead to college credit? Isn't that a BIT excessive...?
" Modern game AI concerns itself with AI for video games. "
No. Not AT ALL. Yes, game AI is a small part of it, but modern artificial intelligence has left game AI waaaay behind in the dust. Game AI is mostly about specialized logic to the rules of the game (pathing, observing enemy moves, etc.), maintaining a priority queue of actions, and making the right responses when the right action is chosen. Out of the 4 artificial intelligence courses I took at college, one of which was graduate-level, we spoke about video game AI for a whole... never. No, we worked on gradient ascent algorithms, simulated annealing, hidden markov models, supervised machine learning, perceptrons, neural nets, but not so much game AI. Chess AI tends to be mostly just calculate all the possible board positions resulting from a given choice, and then the results of choices from those positions, and on and on. Its mostly a brute force problem, our hardware these days can just crunch the numbers. Go is much more challenging for real AI, and thats why we stink at making computers that can play it. If you're talking about videogame AI, thats really pretty simple and isn't AI as we refer to it in computer science. They're still pretty much the same as the FPS bots from Quake 3 or the RTS bots from starcraft. Not a ton of advancement has been done.
That said, the Berkeley Overmind starcraft AI team was pretty impressive, but just to show you how seperate game AI is from real AI, the Berkley team found developing a true starcraft playing AI to be beyond infeasible. So they dedicated themselves (months of development, mind you) to just building an AI that could rush zerg to mutalisks, and then mass mutalisks. Mutalisks are better at responding to tons of microinstructions, they can fly and they have ranged attacks, so a computer can better take advantage of them than say, a melee unit. But you see, the official Berkeley AI team couldn't even begin to handle worrying about build order, different strategies, the game is already insanely complicated. They had a hard enough time just scouting for enemy expansions and so on. And yet, Blizzard included a game AI that can play all factions and uses different unit types. Is it because Blizzard has a FAR better AI development team than the Berkeley research department? No. Its because the Starcraft 2 AI isn't really AI at all, its very specialized case logic. It doesn't learn, it doesn't adapt.
Oh, and it cheats, too. When you turn up the difficulty, they couldn't actually make the AI much better, so they just make it so that insane level computers get more resources than you do. That way its artificially stronger, but not any smarter.
GCS/MU/P d- s:- a-- C++++$ UL++ P+ L++ E+ W++ N o K- w--- O M+ V- PS+++ PE Y+ PGP t+ 5- X R++ tv+ b++ DI++ D++ G+ e++ h-
I know that when I have been in large classes, the added value of taking the class compared to just learning from a book was much lower than with small class sizes. Stanford students may feel cheated if they think that the online portion of this class takes time and attention away from the Stanford class and students.
If you have any interest in A.I. you should check this out. These two guys are legendary in the A.I. world and they are not even dead! Among AI-studens Norvigs book is referred to as 'the bible'. Thrun did more make self-driving cars a reality than anybody else because he is not just very smart but also very charismatic.
The same brilliant AIs that will recognize the worthy questions from students.
Those who would give up essential liberty to purchase a little temporary safety, deserve neither liberty nor safety.
they are the first that does it just like the normal course - with graded work - and completion - for free. all the others you list are just references and the material.
'...if only "Jumping to a Conclusion" was an event in the Olympics.'
i already stuck the syllabus on my calendar - i don't take classes for "credit" i take them to actually learn something interesting and/or useful to me. this is a great opportunity, that may not present it's self again.
'...if only "Jumping to a Conclusion" was an event in the Olympics.'
If it does a job better than a human, then it is an advance.
I think that's what scares a lot of people.
"From the depths of my skeptical and rationalist soul, I ask the Lord to protect me from California touchie-feeliedom."
Can you really earn college credits by taking course on Al Gore?
Sounds like a politically oriented school.
Who would win this election: Andrew Weiner vs Andrew Weiner's weiner.
This is a shameless plug for their own book. (required course material) Is this legal? At any rate, if this is the ethical standard exhibited by their professors, then how can I how can I in good faith hire their students?
Yeah. $103.49 for a text file.
For me, the graded work and scheduling is the key. It's the first free online course that includes a strategy to prevent procrastination!
"[Regarding the 'cloud,'] ownership was what made America different than Russia." -- Woz
not just prevent procrastination but to guide you through the content in a structured and logical manner.
'...if only "Jumping to a Conclusion" was an event in the Olympics.'
I thought it said "Intro to Al" and thought it was a music appreciation course for Weird Al. Stupid lowercase letters that look like capitals...
No, the goal is autonomous systems that can react appropriately to their environment according to their function and task. It has nothing to do with mimicking how a human brain works, although there are sub-fields of AI that do try to work on that.
I'd love to take this course but I know next to nothing about CS. I'm good with computers and I learned python and I can do a lot on the command line -- but anybody can do that. Basically, do people think I can float this, or is a lot of background knowledge necessary?
You're right that most game AI doesn't use very sophisticated techniques, but just for the record it's not true that the Berkeley Overmind team found building a "true AI" (whatever that means) for Starcraft to be "beyond infeasible". They focused on mutas because they're easier to do micro with, but the higher-level strategic code is AFAIK pretty much race-agnostic. There's a build-order planner which can work with any set of building constraints you feed it, there's strategy selection which leverages results of scouting and just needs to know the basic details of units and buildings (e.g. if you observe the enemy building a Stargate, then as long as the AI has been told that Stargates produce air units and that Goliaths (say) can attack air units, then it will shift production towards Goliaths), there's a fair amount of prediction of when/where enemies will expand or attack that has to work for enemies of all races, and so on. The Overmind is certainly not a perfect rational agent (or even a perfect bounded-rational agent, which would be a more realistic goal), but it's much more sophisticated than a bunch of hacks around a mutalisk-micro script.
Source: I'm a Berkeley CS grad student, and I know a bunch of the Overmind authors and have been to a few of their meetings, though I didn't personally contribute code.
It's hard to be religious when certain people are never incinerated by bolts of lightning.
Modern bankers are like that with Java. COBOL is gradually dying as mainframes get replaced, but it's the Java behemoth that replaces it, so six o' one, as the saying goes.
Man who leaps off cliff jumps to conclusion.
I agree in regards to game AI; I said that there was a difference between research and the industry. In the industry it's just finite automata. In research it is neural networks, evolutionary algorithms, reinforcement learning, learning classifier systems, monte carlo techniques et cetera. My own master thesis is based on such AI. The reading I did on the field, also shows that others are researching in AI with these methods.
So yes, we agree. Game AI in research; machine learning and other complex techniques, Game AI in the industry; finite automata and other knowledge based techniques. However, since 2000, research has moved towards video games rather than board games; although research continue on board games of course. Thus the classic game AI, and modern game AI. Modern game AI is game AI applied to video games. For example, the video game I worked on, have its rules procedurally generated, and my AI agents must learn to play it through machine learning; that is a more difficult problem than Go or Chess. It makes sense that game AI should pursue solutions to harder games than Chess and Go, because not only is the challenge bigger, but it also have a potential use in the industry; researchers hope that at some point, the industry will begin using these techniques. It's still not the case however :(
01 REDEFINE REALITY.
Oh, and why it is interesting to pursue AI that can play an unknown game; procedural generated games. You can evolve those, and evaluate them based on how the AI agents play the generated games. If you can make an AI that plays somewhat like a human, you can then in turn procedurally generate games that humans can play. Think of starcraft for example; you could perfectly balance the three races and each unit, through these means. Even adjust the general rules of the game, such as economy and what not. It might show that a third resource is necessary to result in balance. Who knows?
01 REDEFINE REALITY.
Does anyone know whether students have to complete the exams under timed conditions?
"Has the rule of law degenerated into the rule of lawyers?" (Niall Ferguson)
nothing more.
I am very small, utmostly microscopic.