Shared publicly  - 
 
So far, the difference in quality between ml-class.org and ai-class.org is remarkable.

ml-class is exceedingly well done, with fun and interesting programming assignments, lectures that are easy to follow and build properly from easy to harder, critiques of assumptions and common issues with the techniques, quizzes during the lectures that are at the right difficulty and only on material just covered.

ai-class has no programming assignments, no discussion forums, lectures that skip material and jump over key sections (just compare the discussion of linear regression in the two classes), and quizzes that jump from absurdly easy questions to questions on material not covered to unpleasantly time-consuming and error-prone questions (like lengthy probability calculations that require painfully churning through calculations by hand). The production values are also much lower, with bad lighting, no prepared slides, quiz questions that are badly embedded in the video sometimes even covering material in the slide you need to see, and outages on the site.

I hate to criticize classes that are generously being offered for free and involve a lot of hard work from many. But, for those of us taking these classes in part to look at examples of new models of education, I do think it is useful to evaluate these examples and, for that purpose, I think ml-class is by far a better example of how to do this well. If you are taking only ai-class (or started ai-class and gave up), I'd definitely recommend taking a look at ml-class for comparison.
143
37
Daniel Racovitan's profile photoRubén Berenguel's profile photoAttila Babo's profile photoRicardo Colomé's profile photo
68 comments
 
I do find the questions in the ai-class to be more involved and more difficult. The fact that you need to answer a quizz makes you think about the solution instead of just having you apply the theory that came before. However, I do agree with your sentiment. I also like the fact that the slides with notes from ml-class are available, unlike the written shizzle from ai-class.
Yong Yuan
+
1
5
6
5
 
ai-class.org partnered with knowlabs.com, while ml-class.org developed and hosted everything in house. It appears knowlabs.com was not up to the expectation.

Both courses have had light course load so far. The assignments were particularly friendly. Lots of hand-holding in programming exercises. None of the homework questions was like the ones from a traditional university course, which would normally require at least hours of intensive work. I guess the instructors had wider audience in mind. It's a good thing for a diverse audience, and is great for people like me who have a full-time job, but whoever thinks he's being challenged by courses of "Stanford standard" is kidding himself.
 
AI Class is interesting. It's hard because of the instructional incoherence - lectures seems a bit like a journey where all the signposts have been removed. It appears from Units 5-6 that some note has been taken of feedback as these are noticeably easier than units 1-4.

On production values, the presentations are annoying for number of reasons:
- handwritten text is always going to have the instructor's hand obscuring some portion of the material and this makes it hard to pause the video to take notes
- many many mistakes in delivery (spoken and written)
- lack of cross-referencing to the text or other material
- mix of easy quiz questions and really hard ones makes it hard to plan time ( I now find myself devoting 2 hours per unit just to watch it straight through) which is not the desired instructional intention. Easy quiz questions weighted the same as hard ones (same for homework)
- bad pedagogical technique in some places: "Concept X is not important for this class", then proceeds to ask a question about Concept X. Quiz: Can you fit a linear function to this data? Yes/No; next quiz: work out the linear regression. Questions that require answers based on prior experience: "Which one would YOU choose?" marked as Correct/Incorrect when the real question is: "Which one would I choose?" (the instructor) or even the inferred "Which is the correct choice, based on what you know from this material?"
- quiz questions that require spatial reasoning presented with no accurate spatial information

I hear the rebuttal "It's a free class, you get what you pay for. shut up and don't be negative". This isn't valid. Academics are trained to find fault; they have to also be able to accept criticism. Handwaving away flaw doesn't work when there are >100K participants. Moreover, this "experiment" is set up as a beta test for knowlabs.com, which is Thrun's startup. Such criticism is more valid as it will inform the development of the final product. If you are getting your beta testing done for free by a large number of people, you have to take the rough with the smooth.
 
Thanks, Yong, didn't know about Know Labs. Appears that is a very new, very small startup founded just for ai-class by one of Sebastian's PhD students. Here's the former student and now CEO: http://www.linkedin.com/in/davidstavens
 
Whenever I click "Next" during ai-class lectures it never goes to the next thing. I'm using Chrome so I'm not sure if this is the case for any other browsers. Is anyone else having this problem?
 
Also, I unfortunately had to drop the class because of lack of time. That's not their fault though.
 
It could be useful to combine approaches from both classes. AI class is more complex, but sometimes it's hard to understand material from single explanation & quiz... ML's approach for homework is better - I can check that I'm right on different questions & assigments - this helps understand material better. And together with this homework/training, they could keep traditional homework when results are known only after some period...
 
In my opinion it's just two opposite kind of teaching, none of them is superior. While ml class is covering a few subject in greater detail with perfect preparation while the ai class is way broader, that also requires different pace. I tend to learn more at the first place from the style of the ai class, homework statistics shows that clearly, but at the end of the day they are both great!
 
ml-class is better suited ( no paper style) but for the first time online i praise them for making them avail. the best thing would be to use hybrid of khanacademy and mlclass. And always link wiki/khan for underlying or per-requisite issues like probability etc. I think this could be THE important foray of socially acceptable education into the realm of requested free education proliferation
 
Regarding to "... lectures that skip material and jump over key sections (just compare the discussion of linear regression in the two classes)...", ai-class is a general A.I. course, whereas ml-class is fully dedicated to machine learning. You shouldn't blame ai-class for a lack of details on linear regression.
 
Guys, you're missing the whole point. The class project in #ML-class is to write an Octave script capable of watching the #AI-class lectures and making sense of them! Didn't you get the email?
 
Greg, I am taking both the AI and ML class and I totally agree with what you are saying. But I think as far as the material is concerned, it must've been easier for Prof Andrew Ng's team to execute so well because not only ML is a subset of AI but also they are skipping some material that they usually cover in the regular Stanford's AI class. Hence, they had more time to work on actual execution. And using the same argument, Prof Sebastian's team have to cover a lot of material in equal amount of time, so they really can't afford to execute the same way as Prof. Andrew.
 
FYI, Andrew Ng's real class video lectures used to be on iTunes as well a while back at least.
 
I haven't had time to look at the ML class, but, to be honest, AI class isn't hard. While I realize that not everybody can independently learn material without handholding, online education demands a certain level of self-motivation. Mathematical ideas such as simple MLEs and linear regressions are basic topics in statistics. While many schools (including mine) don't require/don't teach at a sufficient level some math concepts such as those mentioned previously, I believe anybody seeking to study AI/ML (and CS in general [as opposed to maybe software engineering]) should have a firm foundation in math.

I feel like Thrun's coverage is fine because he goes over the theory required (especially MLEs and the process [i.e. optimizing the log likelihood function]), but he doesn't require you to know how most of it works/most of the problems follow intuitively. Though, I will admit that the homework questions could afford to be more specific. I got a 57% on the first homework because of an off-by-one error on the graph search problems, though I suppose my HW2 grade makes up for it.
 
I'd have to agree that the #mlclass and #dbclass blow the AI course out of the water. It's not that there isn't a lot of information being presented, but rather that it's being presented in an incoherent manner. There's no integrated way to apply what we have (theoretically) learned. Obviously we can code the algorithms ourselves, but not everyone will do so.

Also, the DB and ML classes both put an emphasis on practical application: "What can you actually do with this stuff? And how do you do it?" I think this is a very important point, because learning for the sake of learning is important, but being able to do something concrete - something practical - with what we learn is equally, if not more, important. Again, students can do this on their own, but integrating it into the class would be beneficial.
 
+Tarik Sturm-Dahal, I do agree that not everybody will implement the methods taught in AI class, but I think those who don't will likely miss out, and in homework 3 I think coding is the best way to get through even the simple examples presented. I'm having a lot of fun coding the Bayesian filter to work with any level of Laplacian smoothing (k>=0) and any number of classes.

+Greg Linden above refers to "lengthy probability calculations that require painfully churning through calculations by hand". Well to answer the questions you have to use a computer, which may also come in handy for performing the tedious calculations! :)
 
I'm liking ai-class, particularly now that I'm writing programs to solve the quizzes rather than hand-calculating things. I would like to take ML afterwards -- don't know how that's going to work once the homework is all over the net....
 
MLClass seems to have had a head start, though. Andrew Ng has been posting lectures and notes up in public for several years now, and even has some earlier public videos that look like a basis for the current class.

http://openclassroom.stanford.edu/MainFolder/CoursePage.php?course=MachineLearning

As far as I know, this is the first time the aiclass profs have published the class for a wider audience.

Hopefully subsequent iterations will be even better for all of the classes.
 
+Arien Malec sure the ML-class programming exercises could be copied, but nobody learns anything by cutting and pasting. The programming scripts run the student's completed code and check its output. This could also be faked in a variety of ways, but knowing the subject being taught is probably the easiest way to get the marks.

The quizzes for ML contain a fair bit of variation and test your understanding of the content quite well.

If students stay honest they're going to get a lot from the class even if the content of the exercises doesn't change from this year to the next.
 
I'm in ml-class, and had planned on taking ai-class subsequently. I've learned a few really powerful things, one of which is that I really dislike Octave (and probably Matlab), and that I'm fluent enough in Python to prototype in it. With all the production that went into ml-class, including boilerplate code and infrastructure, I'm surprised that they couldn't accommodate any other languages. Once I've gotten fluency with the algorithms and tools together, I'd consider offering a class covering the material using Python including scipy and Hadoop or map/reduce and maybe even MacOS Quartz composer and ObjectiveC?
 
Is Octave that bad? For each language they support they will have a new set of potential pitfalls.

The advantage of Octave/MATLAB is that it's a well supported and mature solution to linear algebra and stats, and Prof. Ng can introduce it relatively painlessly knowing that it will work exactly as required.
 
+Tony Sidaway Octave is probably great for people who are either already fluent in it or Matlab or surrounded by others who are (which Andrew Ng purports applies to all of Silicon Valley).

Encouraging the use of matrixes and matrix operations, for example, yielded no parallelism (on my Mac) even though the problems were reduced to "embarrassingly parallel" logic, but they cost me a lot of programmer time. I invested a lot of effort to make my code (literate) match the idioms used to explain the algorithms and the symbolic mathematical expressions so that I could verify my implementation against the spec and play a little more freely with it once I had credit for submitting a correct implementation.

Then again, the concepts are well presented, there's not too much material to handle aside from my job, and I'm probably just learning some stuff I needed, but hadn't resigned myself to yet.
 
I suppose different people will have their preferences. I've written code for a living in a variety of paradigms over the past three decades but I had no problem getting Octave to accept tight, very succinct code and spit out the right answers.

In fact I'm sure the use of Octave saved me a lot of coding effort because it supported the obvious vectorizations so well.
 
Yesterday I got hooked on the ML class videos. I couldn't stop watching them and I had no trouble understanding linear regression with those. Ai class is not that good by far.
 
I don't think it's called for to compare the two courses. For one, they are already both great and generous enough to offer it to everyone. What's better is to help point out how they both can be further improved. Also, reenforcing what's positive in both. ML has a very good pedagogic thoughtfulness in it. AI on the other hand, has been sharing a lot already rich experiences and applications of the subject from the professors. You have a good point in citing how the video production in AI could be better, the exercises, and the forums. I think they are both experimenting on two different interactive learning platforms. In the next iteration, this will surely build upon learning and feedback from all students. Thanks to each classes' team and Stanford's president. They are planning to do more of this onwards.
 
Greg, my sentiments exactly! I'm still doing both but it's a night and day (in favour of ML class) on satisfaction and 'stickability' of learned concepts. More often than not AI lectures frustrate the hell out of me. This was the most visible when going through linear regression bit in AI Class (week after perfecting this concept in ML class) - I was more confused about it after than before pressing play button.
 
+Greg Linden I can't agree with you more. I'm taking both and I find that I'm spending more time on the AI class, being more frustrated with it, and learning less. The section on Bayes was particularly frustrating. Even after doing all those quizzes and homeworks and listening to some lectures several times, I still have no intuition about how various nodes in a Bayes network relate to one another. I got most questions wrong. Afterward I always understood and agreed with the official answer, but couldn't explain why mine was wrong.

In contrast, the ML class is a joy. He talks kinda slowly, but you can download the videos and listen to them at faster speeds most of the time. The website is much better organized. I'm never confused or annoyed.

Also, I don't understand why the AI site is down one or two days per week. They started with 160k students, but surely most of those have dropped out by now (most of my friends have), so you're talking about tens of thousands of hits spread out over several hours, or a few hits per second. How is that taking any site down? All it has to do is check your answers. The ML class has to run your code!

As an aside, I actually like the idea of the periodic quizzes, especially those before you learn the material. They make you think about the question before you're taught the answer. But it's done badly in the AI class. They just make me feel stupid. Maybe if they didn't keep track of the percentage? Or maybe if the instructor didn't act like I should have known the answer.
 
I'm taking the AI class but not the ML. Based on the video and notes from ML, I would say that their approach is more traditional and so familiar to most of us.

As for AI, I believe they are trying out a new approach to delivering courses online. I like the quizzes as it makes the lessons more interactive and allows us to pause and think about what had been taught so far. I think the AI team is probably too ambitious and they are behind schedule. That explains why the video production quality is not up to mark. And since they are delivering these lessons for the first time, probably without any preparation, that explain the mistakes here and there. On the whole, I think the AI team did a fantastic job and I'm sure they will set a new standard and methods of online teaching in future.

I also agree that the course seems incomplete without a project or programming assignment. It would be nice to have a AI game challenge or something like that, with the lectures building up to implementing the project.
 
All good points. I am doing all 3 courses, and although I originally found the concepts in the AI course difficult to understand, I like the fact that they are taking the time to improve. Remember this is an early concept, and I am sure in future iterations of the course they will add more comprehensive material. I don't mind the handwritten notes at all, and it reminds me of some of my lecturers at uni. I agree that more references to the text would be good but overall I am really starting to enjoy the AI course. it seems to be more theoretical than the ML course, which is a lot more hands on. I also like how I could take my program from Octave and run the Linear regression data through it to verify my answers :)
 
you're mostly correct but where I think you are wrong is in your comparison between the two methods of linear regression. In the AI course they spent 10 or 15 minutes going over it quickly. Whereas in the machine learning course we've had 2 or 3 hours of video lectures on it. While the AI course recommends the text book (and I think you need the book to fill the gaps in the course), the ML course is designed to stand on its own without a textbook. The AI class is basically a high level survey across the entire field of AI, whereas the ML class is restricted to a subset of AI so can go into much more detail. Apart from that your points are correct.
 
I'm really enjoying ai-class. The questions take one or two minutes (once understood), and having a quiz every two to three minutes is really excellent for making me pay attention. It forces me to go back and watch the video again if I tune out. In fact, I would say that this class is training me to not tune out!

As +Simplicio Gamboa III said, the anecdotes during the lectures are great motivation. I was really impressed by the body shape/pose demonstration.

I am also okay with the lack of programming assignments. I can practise programming using hundreds of other websites; it's not meant to be a programming class.

Finally, I also really like the handwritten style. I find that polished notes often leave things out, whereas a presenter generating notes as he speaks will often scribble helpful little notes and assistances. It also encourages me to make summaries (as I don't want to have to go hunting back through the slides when I get to a quiz I can't do), a process which is a powerful memory reinforcer.
 
+Francis Fisher I did not realise there was a right and wrong to giving an opinion on the course. I think that comes down to individual tastes and preferences and not someone grading the opinion based on their own view
 
ML class is a ... joke. Prof Andrew Ng has lowered the bar so low that the class can be taken by year 9-10 students. And so far the ML programming assignment is merely rewrite/implement the formulas (from the lecture) in Octave. ML is supposed to be more advance compared to "INTRO to AI", but it turns out the other way around.
 
You mostly summarised my thoughts. I finished ML-Homework+Videos+Anything (9.75/10 in my first try for the review questions, didn't bother to try again) in just short of 2 hours on Saturday. I stomped along the endless flurry of videos and small questions, and did homework for AI in also, around 2 hours. But for ML-class it was mostly fun. I watched a video, went to the review questions and answered them as I learnt the things. In the AI-class, I was asked things I could not answer, was given lousy examples and had a hard fought battle navigating the endless stream of videos. I always end the AI assignments not knowing if I'll get a 30% or a 70%, an unsure of having the right idea in my head... But in the ML-class I end up pretty sure I learnt something and could do it again to completion: I know instantly if I got 100/100 in a programming assignment or not.
 
they do say you have to spend 10+ hrs each week on the AI course, which i guess covers the external reading. Having said that, i totally agree with you that the AI-class does skip quite a lot of concepts and it's quite confusing sometimes. Also i don't know why AI-class doesn't use the same video system as ML-class and DB-class, on which you can speed the lectures up to 1.2x and 1.5x, as well as allowing you to download the videos.
 
There is no doubt that the organization and execution of the ML class is superior. I also find that the AI in-lecture quizzes (which, thankfully, do not count towards your final score) are of inconsistent difficulty, and sometimes seem to assume the very material they are about to teach (or should be teaching).

On the other hand, I'm not as big of a fan of the ML homeworks. I understand the tension: they want to deliver this course to an extremely wide audience with varying programming skills. However, the homeworks are so designed such that you never have to construct a complete program yourself---you just fill in the blanks. I am concerned that some people will develop a false sense of mastery of the material; and then when they attempt to construct an implementation from scratch from their own applications, they will struggle.

But yes, the software platform that the ML and DB classes use is excellent. I am disappointed that the AI team felt they needed to go their own way here; the results are less effective.
 
I think the AI class would benefit from an approach I learned in my physics degree: derive a formula for a given problem, then plug in the numbers and do the calculation once. That would reduce the roundoff errors he gets. The calculator method feels like a high-school approach.

I, too, prefer the handwritten style, as it prompts me to follow along and takes notes as I go. Yes, the quiz problems get involved at times, but the site says the quizzes don't count to the final grade, so go ahead and guess. Usually, I take the time to watch the videos which explain how he got the answer.
 
The professors of both courses said the levels of difficuly of online and on-site courses are the same. Only ai-class has some programming assignments, but that's all regarding the differences.

You shouldn't have inferiority complexes: you are following a fully fledged Stanford-level course.
 
Incidentally, of the three courses, the DB class is far and away the best. It uses the same software platform as the ML class---in fact I believe it was largely developed by the DB class folks, but I could be wrong on that. But the homework assignments in the DB class are extremely cool, because they have embedded relational algebra parsers and database query engines into the online application. So you can build and execute queries right there in the browser-based assignments, iterate until you get the results you want, and submit your solution, with no local software required.
 
+Stefan Park - sorry I was directing my comments at the OP, I did not mean a personal attack on you!
 
Personally, I think the difference in quality is because of one key aspect: Scope. Artificial intelligence as a field is huge. Since machine learning is just a branch of artificial intelligence, it's possible to go into detail about a lot of things (as +Francis Fisher pointed out); but to give a complete course of artificial intelligence in the short span of 10 weeks (I think) only allows for a very shallow survey of each subfield
 
that being said, I do think that the level of difficulty of the quizzes in AI class is... odd... at best
 
sorry but I may not agree with u. I think ai-class is really interesting and encourage students to think over important conceptions by themselves-- by given quizs. I find I understand most of the conceptions after doing the quiz.
Saying 'the discussion of linear regression' in ai-class is not given enough details-- that's not fair at all. machine learning is just a one-week course in ai-class, you cannot expect more.
I think there are too many terminologies in al-class, that's why their lectures seems a little bit hard to understand. maybe we'd better read the book to review some conceptions, Stanford students, they do read books, right?

I agree that ml-class is easier to understand...but I'm not sure if it isn't for the different levels of difficulty of these classes.
 
Either way, it takes a significant extra effort on the part of the teachers to set this thing up and keep it rolling, you've got to give them that.
 
A lot of comments on this post, many of them off topic. The point is that ml-class is a much better example of how to do an online class well than ai-class, so, if your only experience is ai-class and you're interested in seeing a similar but better example of how to do an online class, you might want to take a look at ml-class. Several people commented that db-class is very good as well and worth a look.

A lot of these other comments are beside the point. It doesn't matter that ai-class is more general, that this is Peter and Sebastian's first time, that you think ai-class is easy, that you forgive the problems in ai-class, or that this is a Stanford class. That is all irrelevant if you are looking at good examples of how to do an online class. The point of this post is that ml-class is really well done and, if you are taking ai-class and haven't looked at ml-class, you're missing a much better example of how to do online instruction well.
 
I signed up for ml-class after reading Greg Linden's comment from yesterday. My first impression is that it is better though I wasn't that dissatisfied with ai-class to begin with.
 
I'm taking both, and I do see areas where each has an advantage over the other. The ML class site definitely provides superior options for knowledge consumption, I love that it lets me download the videos as .mp4 files to load and play offline on my tablet. Made great use of my last flight!

That said, personally I like the human delivery of the AI class, it is more conversational and fun for me.

Overall I guess I agree with you +Greg Linden - The AI class can find many areas to improve just by looking over at the ML class. But I don't want the AI class to be done exactly the same way.
 
Did you buy the book? They don't cover things as in depth because it's in the book
 
Disclaimer : I went into AI class knowing I didn't have the prerequisites. I fully realize that AI class would most likely be FAR more rewarding to me if I had those prerequisites. But I REALLY wanted to learn the material, so I decided to try anyway. I realize a lot of people might be in the same boat as me, and get a feeling like the material is too hard. When in reality we'd be able to understand it if we had the prerequisites to the class.

From my point of view the questions before the video's in AI class are way too hard. I feel like he's giving me impossible tasks, and then showing me how to do them after I've failed, this gives me a gut feeling that I can't do it, and it's very discouraging. I can do logical stuff no problem and I always get those right. But with a lot of the slightly difficult ones I tend to just give up, write in 0/1 and look at the explanation instead.

In constrast, in ML class I always try really hard to solve the questions in the quiz'es. If I get something wrong I go back into the material we've gone over and look at it more carefully until I understand it. And with the way you submit programing excersises and homework I feel the focus is on learning, instead of on measuring how well I learn. That said I understand that this might make it slightly harder to grade accurately.

I feel like I can learn without fear of failure in ML class, and that's an incredibly good feeling! Thank you +Andrew Ng!
I've strugled a LOT in the homework for AI class. But I've spent about 20 hours this weekend going over all the material for AI class again. And I have a MUCH better understanding of probabilities and bayes networks now. So I'm not saying AI class is bad in any way, I'm loving it, but it is a LOT harder. But I think if the video's/quiz'es were ordered slightly differently, the class might be far more rewarding to people like me!
 
+Jeremy McMillan
Personally, I really like octave. I can definitely see its potential for prototyping - it is so easy/quick to throw together a script and to monitor what is happening in your neural network.

I've used neural networks in the past: that was really my motivation in taking these two classes as I'm not really a mathematician and have no notation of neural networks that is anything other than lots of linked switches - that makes it incredibly complicated when you're working with 10s of thousands of neurons arranged in all manner of ways, trust me. When Andrew Ng introduced his representation of neural networks using matrices, it was a eureka moment for me. I've used matrices for 3d programming, but never thought of using them in this way. I don't see how you can't see the benefit of this notation. Personally, I hate most maths jargon and tend to be able to conceptulise program flow well enough to not need a mathematical notation, but the neural network matrix notation is super-super important. Andrew mentioned he knew of people in silicon valley making a killing from machine learning without this understanding. At that point I admit to thinking "yeah, I understand them without all that, too. And they're making money, right?" How stupid I was - he wasn't teaching all that math just for the sake of it. It's such a powerful notation - even if I learn nothing more from these classes I will be a happy bunny.
 
First of all to be a part of such a couse as AI is definite skill booster for me. Before this couse there were only two types of problem I faced, one completely solvable two unsolvable. The HW problems require a great lot of deliberation to solve. To understand the problems completly reuires good involvement as they are sometimes between solv and unsolv.

and mostly I am sick of people who are labeling the classes/hw/quizzies too easy, I hardly get a time to take the course and do the homework, for me it's quite difficult. I think the professors are doing pretty good job of keeping the level to the average of the audiance.

They could have easily put in python programming in the online course, but that would have to be something very simple,,,like filling the blanks. AI vs. other classes comparison is unfair beacause of different set of audiances.
Mostly,
DB ML: Programmers and computer oriented ppl
AI: mathematics, concept loving people.
 
Actually both classes have strong and weak points. AI class is frustrating at some points but it's thought-provoking, and it's not a big problem to find complementary material on the web. ML class is very practical and professor Andrew's explanations can be understood even by a monkey but sometimes I have to watch it at 1.5 speed because it's too obvious.

Still it's incredible that we can attend courses in that way. Future has cometh! :-D
 
Funny thing: since my knowledge of math is very limitted, I find the "obviousness" of the ML class teacher very attractive. They actually make feel I DO understand, so this is a driving force to commit to more understanding. On the contrary, when I see other teachers assuming too much and skiping steps or being "clever" I feel like I never gonna get their level. Now: of course I know I could actually go to college, but I mean these are my "gut feelings" about the ways to teach.

I think that teachers should encourage you by any means. I now know that I didn't get any further on math because I lacked good teachers
 
Santosh, if you honestly think ML isn't the most heavily mathematical of all three topics I don't know what to say!
 
+Tony Sidaway Prof. Andrew gives us answers only and constantly stresses that we do not need to know calculus or linear algebra to use ML algorithms. AI course has more math. In AI course they show you derivations and expect more from you and, in my opinion, AI course is definitely more 'mathematical'. Also its quizes which require you to take pen&paper (or racket/octave/...)... Still, I like both courses.
 
Dmitry, I hadn't looked at it that way. Actually I recall that earlier complaints about ML class had centred on the perceived "dumbing down"--Dr Ng had generated linear regression using scalar components. As I've also been following the ML class using his 2008 CS229 lectures I'm probably hitting the mathematical side much harder. Nevertheless the programming exercises are heavily mathematical. Vectorizing the neural network procedures is a rather more heavily mathematical process than anything I've seen on the AI class.
 
I totally agree with all of your points (except maybe with the discussion forums one, because even though ai-class has no forum on the site itself, they do sanction the reddit and aiqus ones).

And I think it is not a bad thing to offer constructive criticism like this.

Both ml-class and ai-class staff have made it clear that these are experiments and first-time tries at a new paradigm in teaching, so things are bound to go wrong on some accounts.
Offering the students' view on what must be improved is the only way they can make it better and I think everyone gains from that.

I am taking both classes and although I had a few notions of AI (from college, 15 years ago) I am truly struggling with ai-class, as opposed to ml-class, a subject about which I had no previous knowledge whatsoever.

But still, kudos to both teams for the effort and a million thanks for the effort.
 
It's hard not to agree with the sentiment that the ML class offers a more polished experience; it also works much better as a stand-alone package, whereas familiarity with material as listed in the readings section and covered in AIMA is often beneficial and, in the case of the prerequisites stated, rather essential for following the AI class productively.

Maybe it wouldn't have been a bad idea not to outsource coverage of the prerequisites to Khan Academy (and stay in-format), or even to start off with some kind of probability/linear algebra self-assessment process. In any case, it would be a pity if frustration with the probability-heavy lessons scares people away from the class altogether.

What I really like about the AI class format is the frequent use of inline mini-quizzes -- a very effective way of keeping minds from wandering (too far).
 
I'm sorry this comment is a bit off-topic, but I was going through the new ones, and I have to agree with +Dmitry Soloviev.

it's just amazing that all these people can receive this kind of education for free! online! it may have its faults, of course, but I'm positive that this is just the beginning.
Future has cometh, indeed :)
 
Greg, the AI class is what I would consider a high-level introduction to AI.

If you were to try and implement any of the constructs that Sebastian covers, you would either have to take a more focussed course or expect to do some research and try to figure out how to apply the techniques to the particular problem domain you are attacking.
 
I understand the sentiments behind this complaint. However, I find myself following the AI lectures much more closely due to the written demonstration. The ML and DB class format feels like a traditional video tutorial even though its not.
 
you cannot compare "linear regression" topic between classes;
aiclass is an introdiction/survey of AI, whereas mlclass treats a small subset of AI, hands-on
 
anyway, I find the teaching style and method of mlclass better than aiclass; I would have liked more interaction from the proffessos, such as office hours
 
but I fing aiclass was more difficult than mlclass because of the proposed topics/concepts in the first place; a lot more new and/or difficult concepts to learn (just think probabilities :) )
Add a comment...