(Re)designing an Introductory AI Course: Post-Semester Analysis

In the Fall of 2015, I was given the opportunity to co-teach Tufts’ Artificial Intelligence class, through Tufts GIFT (Graduate Institute for Teaching) program. As I described in this pre-class blog post, I took the opportunity to re-design the class in order to include more topics from “Modern” (i.e., probabilistic AI). As described in that blog post, I had grand plans, but had no idea to what extent I’d be able to implement them. Let’s look at how Tom’s grand plan worked out!

Outcome

Two things should immediately stand out by comparing the last three columns of the chart above.

First, I was successful in convincing my co-instructor, Prof. Anselm Blumer, to shift most of the sessions on Logic to topics from Modern AI. This was a big surprise for me, as I initially had no idea whether or not he’d be receptive to such a major shift. We ended up dropping HMMs and POMDPs and picking up Factored MDPs, but otherwise the topics covered were just what I’d hoped for.

Second, though, you’ll notice that the actual class schedule jumped around a lot. For example, the last section of the course (16-26) went: Planning, Bayes Nets, Bayesian Inference, Bayesian Inference, Planning, Bayesian Inference, Planning, Markov Logic, Planning, MDPs, MDPs.

This was the product of my planned schedule (ba-dum-psh) rubbing up against my actual schedule. Before the semester started, Anselm and I worked out a rough schedule for the course, and which classes I’d be teaching (I specifically wanted to teach CSPs, Bayes Nets, and Approximate Inference (because I had specific in class exercises in mind), and Markov Logic and Factored MDPs (because these are fairly new topics I wanted to season the main course with (more puns! (nested parentheses!)))). However, once the semester actually started, several wrenches were quickly thrown into the mix.

Best Laid Plans

First, I got papers accepted to several conferences. While this was a good thing for my research progress, it meant that I would no longer be at Tufts at the time I was supposed to teach two of my sessions. To avoid changing what sessions I’d be teaching, we had to get creative with the schedule. This involved many small schedule shuffles, such as putting searching with non-deterministic actions off for a month in order to shift things back a week.

Second, to get students ready for the modern AI section, we decided to combine the lectures on Bayes’ Theorem and Utility Theory, and move them earlier into the course as a sort of “preview for things to come”.

Finally, of course, things just didn’t time out exactly as we expected. Some topics took more time than expected, some took less. In order to stay on the safe side, advanced planning topics got put off for a bit, and then, as discussed, ended up getting peppered in between my other lectures, whose dates were set in stone to accommodate my conference travels. It was a whirlwind.

Conclusions…

My first time teaching was certainly a learning experience. If it weren’t for the fact that I was scheduled to teach very specific sessions of the class, the schedule would certainly have been saner; presumably it will be much saner when I end up teaching the course by myself at some other university, on my schedule alone.

Schedule notwithstanding, however, the semester was incredibly rewarding. I had an absolute blast, and my students seemed to enjoy it too. I ended up getting an average rating of 4.63 out of 5.0 on my teaching evaluations, whatever that means, and one student wrote the following, which made me pretty proud:

“I have never had someone put such a visible effort into each lecture and I really appreciated seeing all the time and hard work Tom put into making sure we had the best learning experience possible.”

Of course, my students were significantly less positive about some of the homework assignments I designed. But that’s a story for another blog post…