CMU's Introduction to Machine Learning Course | Hacker News Reader

CMU's Introduction to Machine Learning Course

alex.smola.org

151 pointssreeix38 comments

I just showed this link to three classmates who are currently taking the course, and the common reaction was "It's a trap!"

They haven't been very satisfied with it. It's co-taught by two professors, one who teaches like it's an introduction for people who have never heard of Bayes' theorem and one who teaches like it's a graduate seminar for people who've seen it all before.

As someone who has never heard of Bayes theorem, is this good or bad? So some things would be explained well and other things over my head?

Right. Some things would be explained from the basics, and some topics would be covered by referring you to obscure papers on advanced techniques in machine learning published by the professors.

I took it last semester. Most definitely a trap - Smola is one of those guys who's just too smart to teach. Great material - terrible instruction.

I am currently in this course.

Honestly, this is not a course that I would recommend. The most problematic part of this course is its lack of clear outline. It jumps between different fields of machine learning, which could have fundamentally different focuses and motivations, without illustrating the connections to the students. It talks about Watson-Nadaraya classiﬁer in the second class, then we have two lectures to explain most basic naive bayes algorithm. I just don't get it.

Though it gets me confusing a lot of times, the course is useful in a way that gives me a lot of keywords to search for and read article about.Also the homeworks might be challenging some time, working through them did improve my understanding of something I might think trival before, like the linear regression stuff.

And if you are really interested, I would recommend a book, which covered most of the materials of the course while being much more organized:

http://www-stat.stanford.edu/~tibs/ElemStatLearn/

A refresher in linear algebra will also help~

Prof Yaser S. Abu-Mostafa's Caltech course "Learning from Data" (http://work.caltech.edu/telecourse) is probably the best introductory course for really understanding the physics of how machine learning works.

See Prof's Yaser's 1 min overview: http://www.youtube.com/watch?v=KlP0DpiM7Lw

The "Learning from Data Book" videos are online for free, and the book is on Amazon...

Videos: http://home.caltech.edu/lectures.html

Book: http://www.amazon.com/Learning-From-Data-Yaser-Abu-Mostafa/d...

The course is also availble on EdX: https://www.edx.org/course/caltechx/cs1156x/learning-data/11...

I've got the book. It's a great book, even though the Machine Learning course here at Technion is more Bayesian than AML's seemingly PAC and VC-focused book.

I also recommend videos and lecture notes by Tom Mitchell for the same course taught in 2011[1]. His explanation to some non-trivial theoretical concepts of ML is very coherent.

[1] https://www.cs.cmu.edu/~tom/10701_sp11/

I did the Stanford free online one the first time it was offered a year or so back. Was perfect -- didn't move at a blazing pace and was very lean. Great instructor, highly recommended (though I think it may have been absorbed into Coursera?).

Yes, I'm taking it now.

Edit: https://www.coursera.org/course/ml

There's also a more sophisticated course on ML by Hinton: https://www.coursera.org/course/neuralnets Have you tried it as well?

I'm browsed it a bit. I'm hoping they will offer it again.

I've been following the self paced AI class in Udacity https://www.udacity.com/course/cs271

Do you have a link? I would be interested.

http://see.stanford.edu/see/courseInfo.aspx?coll=348ca38a-3a...

Nope, it's still around on youtube and on the Stanford Engineering Everywhere site. The coursera version of the class is much more introductory and skips significant parts of the full Stanford version.

http://www.youtube.com/watch?v=UzxYlbK2c7E http://see.stanford.edu/see/courseInfo.aspx?coll=348ca38a-3a...

If you liked the ML Coursera class, Ng also has an introduction to deep-learning in more or less the same casual explicit style.

[1]: Wiki with code, exercises and explanation

[2]: Video lecture one with a recap on backprop

[3]: Video lecture two on Sparse Auto Encoders

[4]: Handouts

[1]: http://deeplearning.stanford.edu/wiki/index.php/UFLDL_Tutori...

[2]: http://www.stanford.edu/class/cs294a/video1.html

[3]: http://www.stanford.edu/class/cs294a/video2.html

[4]: http://www.stanford.edu/class/cs294a/handouts.html

awesome he is a real pleasant & organized teacher for this kindof stuff I will definitely take a look

For comparison, here is MIT 9.520:

Statistical Learning Theory and Applications http://www.mit.edu/~9.520/

I think it is interesting how different ML courses can have such different emphasis in content. The MIT course is all about regularization.

Part of the course requirement seems to be improving the quality of relevant Wikipedia articles: http://www.mit.edu/~9.520/fall13/projects/Projects2013.pdf

I'm taking this class right now and it is certainly an interesting twist on the topic - its fun to see how many different ML techniques solve variants of a single base problem that you can analyze with statistical learning theory. Also: how many different regularizations are equivalent, and how some "intuitive", ad-hoc-seem-to-work regularizations you might think up in isolation actually can be theoretically justified. It contrasts with the more traditional, also grad-level 6.867 ML class.

You may want to look at the self-evaluation, a qualification exam for understanding the material:

http://alex.smola.org/teaching/cmu2013-10-701x/slides/Intro_...

It would have been a piece of cake in college, but I have not used that math in a long time :-/

Gee, guys, looking at the list of topics, a huge fraction, likely over 50%, of the material goes back to programs in operations research, statistics, and the mathematical sciences from about 1970 on. Nearly the only thing new is the collection of sample applications. From what I've seen, the quality of the content of the current ML courses is way below that going back to 1970.

Warning: History shows that the US economy looked at the material in operations research, statistics, and the mathematical sciences and rolled their eyes, did a big upchuck, laughed, turned, and walked away. One might look for alarms from their hype and fad detectors.

a good machine learning course might indeed cover 1940-1980s operations research (nonlinear optimization, linear/quadratic programming, dynamic programming), and statistics from 1970-1990s (graphical models, markov chain monte carlo methods, measures of model capacity). i'd say the field borrows the most useful bits from these fields and finds good honest use in many real life problems today. and i agree that there's a lot of unwarranted hype that leads to a lot of well-deserved skepticism.

That is not the intro course. This is http://curtis.ml.cmu.edu/w/courses/index.php/Machine_Learnin...

How is this course compared to Andrew Ng's Coursera class, his regular Stanford class and Caltech's Learning from Data course? (Other ML courses available on the web in terms of depth)

I can't say how this specific CMU course compares with Ng's online course, but I can comment on Ng's course as compared to several other similar courses I took in Nanyang Tech. U. and Paul Sabatier (Toulouse 3), and my overall remark is that Ng's course is quite short on the maths, which makes it not sufficiently formal to deeply understand what goes on. However, it gives enough material and code samples to play with data. It can be nicely complemented with some self-study.

It's worth clarifying that Ng's online course is essentially his Stanford course minus most of the math. The online version doesn't have the proofs or theory problem sets, while the Stanford version does. The problem sets, etc... are available at cs229.stanford.edu, if anyone is interested.

I will say that I'm a huge fan of the Caltech Learning from Data course (currently also offered on EdX). I took Andrew Ng's Coursera course 2 years ago, finished it successfully, and liked it. But I feel that the Caltech course gave me a much deeper foundational understanding of the basic issues and tradeoffs, and much deeper insight into what's going on.

Homework is much better in the Caltech course, too. In the Coursera course, they give you programs and environments in Octave that are all prewritten for you, and you just need to plug in a few key lines (often there's essentially one way to do it due to dimensionality). You feel like you understand what's going on, but the understanding is not really grounded. The Caltech course has multiple choice questions, but they look like this: "implement this algorithm, run it through a data set chosen randomly with such and such parameters, calculate learning error, do all this 1000 times and average. What value out of these 5 is your learning error closest to?". You choose the language, you implement the algorithm from scratch, you debug the hell out of it, you visualize your data to understand what's wrong... then the knowledge and the understanding stay with you.

I'll second that.

I'm currently doing both the Coursera course and the Caltech course concurrently. I really like the level and delivery style of the Caltech course. It covers a lot of material, with good depth and rigour where needed and with a lot of colour. Makes you want to jump and try the techniques out.

In contrast the Coursera course seems a bit easy and dry. I also dislike the dependency on Octave.

I think now it is the time we get some tutorial/resources/classes on practical implementation of these ML techniques. Enough of Introduction to ML. How to handle large data (say 6000000 rows), how to convert csv/tbv data to different formats needed for different machine learning libraries for e.g. Weka, LibSVM etc.

For an introduction to the broader realm of data input, normalization, modeling, and visualization -- in which ML plays but a part -- you can "preview" Bill Howe's "Introduction to Data Science" class on Coursera[0]; I'm working through the lectures, and I find he gives compelling explanations of what all these parts are, why they're important, and how it all fits together in a larger context.

[0] https://www.coursera.org/course/datasci

I took Prof. Howe's course on Coursera and it's a bit of a mixed bag. I can actually see it being better in some respects just going through the content after-the-fact than taking the course as it was run as there were a number of issues with auto-grading of assignments and some of the specific tools choices (like Tableau, which only runs on Windows).

That said, the course covered a lot of ground and touched on a number of different interesting/important topics. Some of the lecture material was a bit disorganized/had errors and didn't flow all that well from one topic to another but there was a lot of good material there, especially if you had enough background to appreciate it. I was comfortable enough but it was obvious that the expectations set by the prereqs were off.

Hopefully the course will run again with most of the kinks worked out and, perhaps, a better level-setting of what's needed to get the most out of the course.

Here's the averaged perceptron used in a part-of-speech tagger: http://honnibal.wordpress.com/2013/09/11/a-good-part-of-spee...

Because the learner is quite a good fit for the task, it performs better in terms of speed/accuracy trade-off than many other algorithms, such as CRF.

A follow up post for statistical dependency parsing should be finished in about a month (it's down my queue...)

It's worth noting that very many CMU courses, from all departments, have notes available on a publiclly-accessable course website.

Source: Am a CMU 2012 grad.

I think you meant noting* ... haha

All I can say:

1971–75: DARPA's frustration with the Speech Understanding Research program at Carnegie Mellon University

Unless it is blackboard with chalk, its no fun following online. That's why MIT's lectures rock!

Yes! I've done quite a few online courses now and the ones I've enjoyed the most have all been ones that were essentially recordings of the normal classes students take on campus.

In the case of Harvard's CS50x, it was essentially the exact same course. (Plus it helped that David Malan is an outstanding teacher.)

Reply on news.ycombinator.com