Background
I first took CS3600: Introduction to Artificial Intelligence in Fall 2015. In Fall 2016, I decided I wanted to TA for the course, and that was my first term as a TA. Being a new TA, I was still getting the hang of the material, being able to explain content in my own words, and help student debug all sorts of issues.
In Spring 2019, during my Master’s program, I decided to TA for the course again. This time, I had graduate-level experience in some of these topics such as Machine Learning, so I went in with much more confidence ready to help students grasp the material.
Finally, in Fall 2019, I was chosen to be the Head TA for the course, which now had two sections with a total of 450+ students and 18 TAs.
Mass Autograder System
During my term as a Head TA, one of the biggest challenges I faced was the issue of grading projects. With the number of students we had and the pace of the course, we simply couldn’t expect that all TAs would be able to grade a large block of students manually with only a few days notice, nor could we expect students to wait 3+ weeks to get feedback on their submissions.
Although the projects we used for the course already came with test cases for students to validate their implementations, as TAs we still had to run our own tests to ensure that students didn’t trivialize the assignment by using functions or libraries we explicitly forbid. Additionally, we also wanted to run plagiarism detectors to ensure students didn’t use code they found off of the internet.
The previous solution we had for doing this in bits in pieces had a number of major issues:
- Brittle: it was unable to handle submissions that weren’t in the exact submission file hierarchy the assignment requested, even if all of the necessary source files were submitted.
- Unreliable: Sometimes the library would just not be able to handle common compression formats like _gzip_s, and these faulty submissions would have to be identified and regraded manually.
- Lacking key features: Graded submissions would be output in raw text files, meaning a TA would have to manually go in an enter those grades in the gradebook, all 450+ of them.
I decided it would be worth the time to create a new solution that would remedy all of these issues and would be easy enough for future Head TAs to run without any issues. Here are some of the key highlights of the implementation I completed the first three weeks into the semester:
- Automatic exporting of grades to the full gradebook
- Support for submissions with variations on the file hierarchies
- Easy extensible for future projects to be added to the course
- Improved plagiarism detection
- Extensive logging to ensure TAs can catch issues early on
Project Review Sessions
Although this course doesn’t have formal recitations, the TAs for this course have typically taken it upon themselves to lead several review sessions throughout the semester for students to hear the material delivered at a slower pace by their peers as well as ask any questions they may have on course concepts. These sessions were mainly geared towards preparing students for the upcoming projects since they were such a large portion of the students’ grades.
In previous semesters, the TAs had typically explained concepts verbally, with a bit of psuedocode given for common algorithms such as Djikstra’s Algorithm or the Bellman Update Equation, however I had always felt as though students were able to implement these in the projects and get by in the course without having a firm understanding of why they worked. Many students struggled heavily on the exams in those same concepts that they glossed over since they could implement the projects without that understanding.
To remedy this, I decided to create extensive review guides for the review sessions that could be used long after I had left Georgia Tech. With a handful of other TAs, we managed to produce well over 30 pages total of explanations of core concepts for the course.
Here are the review guides we created:
- Graph Search Algorithms
- Reinforcement Learning
- Probabilitic Inference
- Neural Networks
- Decision Trees
In addition, these TAs and I led the hour-long review sessions for the projects where we presented this material to our students. Sessions typically had 30-40 students, and we addressed questions with illustrations, equations, or any other means of getting the students to understand the material.
Reflections
It was a true honor and privilege to be the Head TA for this course. Not only did having so many students push me to refine my own understanding of even the subtle edge cases, but it also gave me the opportunity to bring what I have always felt were some very necessary changes to the course. I’m hopeful that the students were able to reap all of the benefits of these changes and that it truly inspires them to pursue research and careers in Artificial Intelligence.