Initial Thoughts on (Re)designing an Introductory AI Course

A month or so ago, I was accepted to the Tufts GIFT (Graduate Institute for Teaching) program, which is comprised of (1) a learning portion in which I’ll spend three weeks taking classes on teaching pedagogy, (2) a design portion, in which I’ll be working with Professor Anselm Blumer to redesign portions of Tufts’ “Introduction to Artificial Intelligence” course, and (3) a teaching portion, in which I’ll teach some number of lectures (I’m hoping eight) for the redesigned course.

While I haven’t yet undergone the learning portion yet, I’ve been eager to start brainstorming how I might change the class. One complaint many (including myself) have with how the course is currently structured is that it is has very little of what would be considered “Modern” AI. In fact, 40% of the class is focused on logic, which, as I’ll show later, is quite a lot. This led me to formulate my first goal:

Classical and Modern AI should be given approximately equal weight.

Another complaint many (including myself) have with how the course is currently structured is that the course currently seems to just follow the progression of the textbook, chapter by chapter, without sufficient motivation for each new topic. This led me to formulate my second goal:

Each topic should build off the previous topic in some meaningful way. When discussing each new topic, students should have to consider how it compares to the previous topics studied.

By thinking about the course in light of these two goals, I realized that the course material could be nicely organized in the following way:

• First, there are a set of foundational topics (e.g., Search and Constraint Satisfaction) that (I believe) sit outside of the Classical/Modern divide
• Second, all the topics I’d like to cover in Classical and Modern AI can be categorized as knowledge representation, inference, or decision making.

By adding seven classes to the schedule on modern inference and decision making, cutting six classes on classical knowledge representation and decision making, and cutting a single class on Foundations, I constructed the following schedule, which achieves my two goals. I’m especially excited about the lesson on Markov Logic Nets, a topic which beautifully ties together Classical and Modern AI.

Comparison to Other Schools

The first question you might ask is whether my proposed cuts are justified. To examine this, I looked at a random sampling of universities whose syllabi I could easily find, and looked at the percentage of the semester each university spent on Foundations (Search and CSPs), Classical AI (Planning and Logic), and Modern AI (Probabilistic Graphical Models and Markov Decision Processes).

This showed that Tufts currently spends slightly more time than average on Foundations, four times as much time on Classical AI, and 40% as much time on Modern AI. You’ll notice that percentages for most schools falls far short of 100%. This is mainly because most other schools spend a portion of their AI class on Machine Learning, another facet of Modern AI, which we’re explicitly looking to avoid since we have an entire course devoted to the basics of Machine Learning. I think it’d be fair to say that we thus spend closer to 20% as much time as other schools on Modern AI once Machine Learning is taken into account.

Outlook

How much I’ll actually be able to modify the course syllabus is still up in the air, but I’m hoping I’ll be able to bring it significantly closer to my ideal. I’m guessing that MDPs, POMDPs, HMMs and Hierarchical planning might get the axe in order to add back some of my cuts to the (egregiously long, in my opinion) section on First Order Logic. Only time will tell. As this syllabus evolves I’ll post regular updates. And, as always, I’d welcome feedback on this course layout.