I’ve spent the past 9 weeks (i.e. weekends) taking Caltech’s Learning from Data online class. Each week I watched two hours of lectures and did the homework, which was a multi-choice quiz that typically required writing some machine learning code. Overall, it was a great experience and I’m glad that I took the class.
I learned a lot about machine learning algorithms; got to exercise my very rusty math skills since a lot of machine learning involves statistics, calculus, and linear algebra; had a hilarious time trying to explain a gradient descent algorithm to my family; and had the experience of having 3 hours before the homework deadline to learn enough Python (a language I’ve avoided until now) in order to use a nifty quadratic programming package.
This was my first online class. I’ve come to realize that despite all the hype, online classes have a mix of benefits and drawbacks. On the one hand, I liked the convenience and low cost (ie. free) of the online class. It’s unlikely that I would have found the time to do a real world class. I also learned a lot from the class so that was definitely a good thing.
On the other hand, the format of the class could have been better. Each week, I had one shot to take the test. This mean that sometimes I lost 30% because of some stupid error:
“what you mean excel’s log() function is base 10, even though there is a log10() function too!”
“when computing support vectors, anything < 1e-05 should be considered zero. Seriously?!”
“So the identity matrix only has 1′s on the diagonal. Doh” (Told you my math skills were rusty )(Note to prospective employers: this was feedback from a friend, not from me )
It would have been less frustrating to have had either test data to validate the code or multiple quiz’s.
It also would have helped if I’d had a real human being to talk to about. You know, like in those totally obsolete, bricks and mortar colleges. The class does have a forum, which was helpful, but 1-on-1 time with a tutor is priceless. Coincidentally, I’m currently reading To Save Everything, Click Here: The Folly of Technological Solutionism, which is an incredible book. It references a study that found that the best predictor of student success was the amount of interaction they had with professors. Of course, that costs money and doesn’t scale to 50K person-sized classes, which I think the online class fans will need to accept at some point.
So ironically, there were some real challenges for human learning in the machine learning class. But despite those challenges I’m glad that I did it and would highly recommend the class. Another session is starting in April.