Stanford University's great experiment with an online AI course has reached the midterms and we thought it was time to turn the tables and submit a report on how it was all going. Is it living up to the promise? What are the highs and lows? How does the technology perform?

When we reported the news about Stanford's free online courses for Fall 2011, a number of I Programmer contributors were really excited by the idea of taking a top notch university course on AI and be taught by Professors Sebastian Thrun and Peter Norvig.

Thrun and Norvig's experiment of offering Introduction to Arificial Intelligence as a free online AI course, complete with online exams, attracted huge student numbers. Just prior to the start of the class Peter Norvig posted on his Google +1 account: There are now over 140,000 of you. And three of the number were members of the I Programmer team - Mike James, Harry Fairhead and Sue Gee.

This report has been compiled by the three of us. And just as Peter and Sebastian take it in turns to teach the class we'll do the same. As students we have pretty different backgrounds and different reasons for signing up. 

Mike had completed a PhD in AI, on reinforcement learning which turns out a big part of this course, a few years ago and wanted to do it as a refresher.

Harry has been involved in robotics and the hardware side of AI for a long time and wanted to add some of the mathematical theory. Due to pressure of work he decided to do the "basic" course without the exams.

Sue on the other hand had just completed a course in robotics and wanted to go further and really hoped that at the end of the course she would have merited the signed certificate.

Harry's take - the technology

From the word go the first shock was the very basic nature of the technology used - You Tube hosted videos of talking heads and, mainly, disembodied hands writing in real time on scraps of paper - sometimes with less that perfect handwriting. Needless to say there are mistakes which have to be corrected with small notes below the video.

This is not polished!

In a world of slick "PowerPoint" presentations and animated diagrams  this looked all a little home spun.

And then there was the disastrous attempt at holding office hours using a Google+ Hangout - something I thought was going to be a revelatory experience. It just didn't happen.

Sebastian and Peter's faith in Google just proved misplaced!

As it turns out, however, the personality of the two profs comes through often enough to make it work and having the videos broken down into small chunks works fairly well for those of us who are trying to fit the course into the cracks.

If you have a less-than-perfect Internet connection, it can be frustrating - but there are simple remedies such as getting the whole clip to download before starting to view.

Technology also makes the course more accessible. It's often more convenient, for me at any rate, to read the caption text than listen to the audio, thereby eliminating the scratching sounds made by pen on paper and using the translation facility the captions can be shown in a wide range of languages. 

Another clever part is the use of selection boxes and text boxes to allow the student to enter answers to quizzes that were more than simple multiple choice questions and which could cope with a range of correct answer. So for example it recognized, say 2/3 and 0.67 and 0.66666 as being equally valid answers - something you might take for granted but I have used courseware that was far less tolerant.

Mike's take - the course content

We are now over halfway through and so far the course has largely been about statistics and probability with a little logic thrown in. True we've been told about how this applies to real situations in artificial intelligence, particularly in robotics and with the specific example of Sebastian's work with the Google self-driving car. 

As to the course content the a common problem was the lack of a complete explanation of an idea. For example in the context of Bayes Networks the concept of d separation was introduces without an accurate definition. Students were expected to absorb the idea by looking at examples. So the obvious thing to do is to look it up in the course text book - but it isn't there. Looking it up on the web provides very little information and what there was used different notation, jargon and might not even apply to Bayes Networks but causal networks instead!

There is no doubt that the course is sloppy, vague, inaccurate and just not sufficient to get you up to speed on any of the topics unless you already have a reasonable grounding in the maths. For example, after a few seconds of introducing probability theory the course is working out results that rely on conditional probabilities. This is tough for the beginner.

The same for logic. A few minutes of propositional logic, then first order logic and the student is expected to work out truth tables or reduce formulae to discover if they are valid etc. It is all too quick. You can't master probability or logic without doing more practice than the course provides. You have to find other resources and spend more time getting up to speed.

Worse is the often very vague and loose method of presentation and question setting. In the absence of a precise definition of terminology or algorithms you are left guessing at what a question might mean. There were even examples of the midterm being corrected as it was being taken by the thousands of students.

But before you go away with the idea that I'm a disgruntled student - far from it.

If I wasn't enjoying it I would have dropped out - that is one of the great benefits of a free course. While I can see its shortcomings, this is enough fun for me to devote a large proportion of my free time, and also quite a bit of time I've had to beg, borrow and steal, but don't tell anyone..