I've been taking some online courses through Coursera, a co-operation between Stanford and some other places. I originally signed up for 'Model thinking' (a rag bag of quantitative and semi-quantitative models for social science), 'Game theory' and 'Information theory'. The latter shows one particular problem: The course never happened. The other two merely had delayed starts. The other problem is that the courses tend to be very much introductory-level. They assume nothing, and tend to let mathematical detail slide. Don't pretend for one moment that you're attending Stanford. I guess they're pretty accessible, though.
The content is generally pretty good, there are review quizzes and the like, and one of my favourite features is that you can watch the course video at up to two times normal speed. It turns out I can listen a lot faster than people like to speak, especially if I can pause the video for a good think should something subtle or clever appear.
The 'Model thinking' course was mathematically very simple, but it opened up a new way of thinking to me. I'm used to the physicits 'Put it all together into a unifying framework' approach, with models that can be complex, but are (mostly) testable. Social science lacks a grand unifying theory, but instead you can make do with very simple models to get a qualitative feel for behaviour. When you have a problem, throw as many different models as you can at it, and see how they agree.
'Game theory', as an introductory course, didn't feel like it taught me much I didn't more-or-less know already, given I'd read a not-very-good text on the subject a few years ago. On the other hand, it pressed the knowledge home, through explanation on videa, plus a variety of quizzes. It demonstrates the strengths and weaknesses compared to, say, just buying yourself a big textbook and reading it. The depth won't be as great and/or the pace will be slower, but it seems to me to be a much more effective way of jamming knowledge into your head, even if there isn't personal feedback from a tutor.
I'm now working through 'Machine learning'. My view of AI is... not wholly positive, and I'd kinda got the feeling of this kind of stuff as people farting around and not really getting anywhere. However, in the last decade, it's all started to click. I mean, Google's doing translation using machine learning and statistics, cheapo cameras have face recognition built in, and all this kind of stuff. Huge data sets and huge computing power, plus new research have made all kinds of stuff possible.
'Machine learning' isn't trying to screw around passing the Turing test, either. What I realised when looking at this course is that it's basically fancy forms of regression and classification. Indeed, the course starts off with linear regression! In a sense, it's all super-fancy curve fitting. There are programming exercises, including cool stuff like using neural networks for OCR.
As I said, though, these courses tend to be rather introductory level. The pacing of the course and exercises means that you've got a better chance of absorbing the basics than just picking up a big textbook, but I don't think it really suffices. So, I have bought a big book, namely Bishop's Pattern Recognition and Machine Learning. I'll see how I get on with that.
And there are plenty of courses I'd still like to do. I'm having to ration myself to prevent either me or Caroline from going mad, but it'd be nice to do the machine vision/image processing course...