CSCI 2951X: Reintegrating AI
The primary goal of Artificial Intelligence has always been to build complete intelligent agents. However, the field has also always been fragmented into a collection of problem-specific areas of study. This seminar course will survey efforts made, over several decades, to produce "big picture" theories and architectures for reintegrating the various component technologies into complete, generally-capable, intelligent agents. The class will read and discuss two papers per week. Grading will be based on two written essays, and a substantial open-ended final project.
The first class is on Thursday January 25th. The class meets on a Tuesday-Thursday schedule, from 1:00pm to 2:20pm in CIT 506.
Note that the schedule below is tentative, and may be revised as we go along.
GOFAI and connectionism are generally conceived of as direct competitors. However, their relationship is not quite that simple, since they are usually not applicable to the same sorts of problems. I'd like you to think about a specific type of problem - two player, zero-sum games, like chess and Go - where both approaches have been tried.
Your aim is to discuss the relative merits, successes, and failures of both GOFAI-like (which we'll take to include all search, knowledge-based, and explicit reasoning systems) and connectionist (which we'll take to include the direct prediction of best move or value using a neural net, of whatever type) systems. Where both types of methods have been combined in a single system, I'd like you to explain why, and to analyze what specific advantages the combination brings to the game. Your goal is to try and characterize what specific aspects of a two-player game GOFAI-style and connectionist approaches are best suited to, or to demonstrate that one approach has decisively "won".
I expect you to do substantial reading outside of the course materials, and to write a properly referenced report. You should cover as many individual games as you feel is necessary to cover the "space" of solutions. (I would be surprised if that is less than three.)
The assignment is due in class, in hardcopy, on March 15th. It may not be more than 8 pages in 11 point font (not including references). Please do NOT feel the need to necessarily use all of those pages; I am grading on insight, analysis, and coverage, NOT length.
Your second assignment is to compare connectionist and Bayesian approaches to Natural Language Processing.
Here I'd like you to pick a component problem in NLP (e.g., parsing), and find and summarize the state-of-the-art approaches that use deep nets and that use explicit probabilistic models. I'd like you to discuss both the actual performance achieved, and the specific things that each approach makes possible in principle (e.g., semantically meaningful uncertainty) for the specific task you've chosen.
If the leading approaches are hybrid, that's OK, and even interesting! Explain what advantages are imported from each paradigm.
I expect a properly referenced hardcopy report handed in during class on April 17th. It may not be more than 5 pages in 11 point font (not including references). Please do NOT feel the need to necessarily use all of those pages; as usual I am grading on insight, and analysis, NOT length.Final Project
Your major project will be a substantial creative and original piece of work, in one of the following two categories:
1) A computational research study, in which I expect you to design and implement theory and experiments studying some algorithm or model that is aligned with the course content. I expect it to have approximately the length, form, and content of a conference paper at a good AI conference (though it need not be publishable).
It does not necessarily have to be an original contribution, although that would be nice. If it is not, then I expect at least an original evaluation that is relevant to our topic (e.g., an existing algorithm is tried in a new domain that speaks to what we have discussed).
For example, one might design a new method that combines connectionism and probabilistic reasoning, or one might find such a method and then evaluate it in a domain designed to test one of the critiques of such methods that we have discussed.
The study will be graded on insight, completeness, and clarity. These studies can be completed singly or in pairs. I expect the submitted document to be more substantial for a pair than a similar single-author assignment that earns a similar grade.
2) A philosophical study, in which you propose and defend a proposition about the nature of intelligence. Here I am looking for the sort of argument and analysis that we have been seeing in the papers we've read so far. I expect a well-referenced, thoughtful piece that advanced a coherent and interesting argument and interrogates its implications.
For example, one might propose that (following today's discussion) a text corpus, no matter how large, does not contain the information necessary to discover the meaning of any sentence in natural language. A good essay will describe the proposition, argue for its truth, point to (or argue the irrelevance of) results in the CS literature, and perhaps consider and argue against possible objections.
Again this does not have to be a particularly new idea, but I'd like you you to argue it without relying on the arguments of others. You should cite computational research studies and foundational philosophical studies (e.g., Chomsky), but do not find another essay on a similar topic and use its arguments.
This type of study will be graded on the coherence of the idea, how well it is argued and analyzed, and on writing style. I expect it to be about as long as it needs to be to be complete; 10-20 pages as a ballpark.
The project accounts for 60% of your grade and is due at the end of the reading period (May 8th).
Graded components will tentatively include two written homework assignments (20% each), and a substantial final project (60%).
Our readings are in large part drawn from the following books, all of which are highly recommended for more in-depth reading into this topic:
These books and readings will be of interest to students who want to understand the probabilistic foundation of AI more deeply: