Knowledge-Based AI

I don't really get all the complaints about this course. Some on reddit said "this is the worse class ever" while others on discord were saying that it was BS...uh...were we taking the same class? I guess one valid complaint is that the lessons aren't very applicable to industry, but like, that describes 70% of the content in this master's degree (except for GIOS, that class was amazing).

I thought the class was pretty decent, even good at times, actually. Like sure, it's a David Joyner class so it will have a fair bit around of busy work but overall expectations were very fair and the homeworks and projects were pretty relevant and reasonable. I thought the course was altogether quite easy especially compared to summer ML4T and the TAs were very helpful.

I wasn't a big fan of the lectures but some of the projects were fun and relevant. I personally reused my decision tree code from ML4T for one of the miniprojects and you got introductions into important algorithms like A*, DFS, BFS if you chose to use them for the miniprojects. The new ARC-AGI project was also pretty fun and while some say it was harder than the old raven's projects, I thought it was very reasonable so long as you don't procrastinate. One key thing to note is that you get points for the training problems too, so literally you are guaranteed at least 50% of the code portion of the points if you aren't lazy.

Some complained about the TA grading but I didn't have any issues and got literally a 100% for every written assignment I submitted. All I did was open the rubric (there should be a table rubric for each assignment) and I wrote to the rubric. Put my subtitles in my paper as the exact same row name in the rubric and wrote to the rubric, and I never had a problem with grading. Heck, I thought the grading was more lenient than most of my undergraduate classes. Seriously y'all. Were we even taking the same class?

Overall, 10 to 15 hours a week for this class as long as you are consistent. Pretty moderately easy. I think hardest parts were a few of the mini-projects and milestone D of the ARC-AGI projects, but if you do well enough on the mini-projects you can afford to half-heartedly address milestone D.

This course was my first OMSCS class. I came away with mostly positive feelings about the class.

For starters - Dr. Joyner and the TA we excellent. They were super responsive, engaged, and enthusiastic. I think that the decision to migrate the semester project to ARC-AGI from Raven's Progressive Matrices was awesome. I appreciate how much work that must have taken - and I feel that it greatly enhanced the learning experience.

Another positive aspect of the course is how well organized it is. It is clear from Day 1 exactly what you need to be successful in the class. Everything is available from the jump and this would be a great class to work ahead in if you wanted to. There are many assignments throughout the semester. While this can be a bit grating, or feel like busywork at times, they were generally interesting. And opportunities for easy points. Sometimes other reviews for other courses discuss being annoyed with the uncertainty - and I feel like this class is the antithesis of that.

The lectures I found interesting at times. Unfortunately, they didn't feel as "rigorous" as I wanted them to. I feel like there are more abstract topics in computer science (like algorithms) are theoretically/mathematically well founded. Or there are more practical classes (like Operating Systems) which are grounded in their practical application in the real world. KBAI feels like a course discussing a particular view of artificial intelligence which is neither mathematically "true" nor broadly in use. It is a very "vibey" class which kinda left me feeling like I was not learning a "real" computer science topic.

Some people in the class complained about the peer review software or some of the other administration of the class - but that never felt like an issue to me. Also - people who were complaining about not getting perfect scores on the assignments seemed to be missing the forest (an A is extremely attainable in this course) through the trees (missing 5% due to poor formatting on a report without comprehensive enough TA explanations).

In summary - I think this is one of the best ran classes I have ever taken. It really does not waste your time teaching you the material along the way. Ultimately - the subject matter / topic did not completely resonant with me. Which colors my rating + feelings of the class. But depending on what you want out of it - it could be a perfect class.

Background: 1st semester student, Undergrad CS major, working as a Data Scientist for less than a year.

Finished the course a month early and ended up with an A(91.89%).

Pros:

• The lectures are very interesting and enjoyable, and they are presented in a way that’s easy to understand.
• The coding assignments aren’t too difficult and are manageable.
• The TA and Dr. Joyner are great and respond quickly when help is needed.
• The integration with gradescope is cool, you are able to submit coding assignments up to 40 times and it auto grades so you can see your grade right away. Cons:
• The lectures don’t align very closely with the assignments; the assignments aren’t really AI focused and feel more like general coding tasks.
• The writing assignments feel somewhat useless and repetitive more like busywork that could be avoided.
• I received feedback on early assignments, but later in the course I didn’t get any feedback at all.

Homework(15% )- 90/100: These are writing assignments (journals) based primarily on lecture material. Make sure to follow the prompt questions closely and answer them clearly and completely. Demonstrate a strong understanding of the relevant lecture concepts and apply them accurately. When a prompt asks for a diagram, ensure that it matches the one shown in the lecture exactly. In a previous assignment, I mixed up two concepts and received no credit for either, which resulted in a C on that assignment so accuracy is very important.

Exams: Midterm(7.5%) - 70.91/100, Final(7.5%) - 92.72/100: I’m not a great test taker, so I didn’t do very well on the exams. They aren’t too hard since they are open notes and open internet including AI, but you still need to understand the concepts really well. On the first exam I struggled with time management and ended up turning to ChatGPT as a last resort, which didn’t help much. After that poor result, I made really good notes and used NotebookLM by Google to ask questions based on them for the second exam, and I ended up doing much better. I recommend that if you are unsure if a statement is correct to leave it unselected.

Mini Projects: Performance(15%) - 98.5/100, Journals(15%) - 95.4/100: The coding assignments are fairly easy and similar to medium-level LeetCode problems, so they shouldn’t take too long, except for Mini-Project 2 (Block World), which was difficult and cost me a few points. Overall, the assignments are manageable. The writing assignments are short (max 4 pages), but I lost points for not including metrics related to efficiency and performance.

ARC-AGI project: Performance(7.5%) - 100/100, Journals(7.5%) - 100/100: For the assignments, you’ll be solving ARC-AGI problems by creating an agent. I didn’t implement any real AI methods myself, what I did was design a specific solution approach for each problem. To achieve full credit on each milestone, you need to pass at least 6 out of 16 general test cases and 6 out of 16 hidden test cases. The journals can be repetitive, so feel free to reuse the same structure each time. Just make sure you include specific metrics for each milestone, including: Efficiency: Big-O complexity and actual runtime, Performance: Number of test cases passed Keep in mind that even though you only need 6/16 for both general and hidden tests at each milestone, you’ll eventually need to solve all of them for the final project. Because of that, I recommend completing as many problems as possible throughout the milestones rather than waiting until the end.

Final ARC-AGI project: Performance(7.5%) - 84.38/100, Journals(7.5%) - 76/100: Each milestone (Milestones B through D) includes 16 general tests and 16 hidden tests, for a total of 48 general tests and 48 hidden tests across all milestones. I wasn’t able to solve 15 of the hidden tests, which left me with a performance score of 81/96. My performance wasn’t as strong as it could’ve been because I decided it wasn’t worthwhile to spend hours trying to figure out each hidden test for just a 1 point increase. The final journal grading is much stricter than the other journals. I answered every question on the assignment, but some parts were considered vague by the TAs, which lowered my score. I put in the same level of effort as I did for every milestone, where I earned 100 out of 100, but the TAs expected more for the final journal. I recommend being very clear and keeping in mind that the grading is stricter.

Participation(10%) - 100 / 100: These points are basically free, and there are plenty of opportunities to earn them. I mainly completed peer reviews each week to stay ahead. After about two months, I had already earned the full 90/90 points and didn’t have to think about it for the rest of the semester.

I found the lectures interesting but not terribly helpful. I really like the professor and admire the effort he puts in to provide a good learning experience and a great class though. The goal of the lectures is to teach first principles to use solving ARC AGI problems, but because there's not a lot of machine learning from data it feels less like an AI course and more like a few dozen leetcode problems. Some concepts were more advanced, like incremental concept learning (decision trees updated one data point at a time) but many AI concepts felt more like I was the only intelligence involved (represent problem with a state space, use guess and check to solve, etc). Most of the hours I spent on the class were on the semester long project, with took me around 100 hours. In retrospect I think it would have been faster and easier to ignore the course concepts and write a separate algorithm for each of the 48 types of puzzles than trying to build a general solver that uses course concepts like I did. The tests weren't fun but they felt pretty pair for other GT classes, and the rest of the mini projects were pretty quick. The course has a significant participation grade that required reviewing other peoples reports (required with each coding project). I wasn't a huge fan of that aspect, and felt like the reviews were not helpful, but they weren't too inconvenient.

TL;DR: - OVerall, I liked the course. There is a lot of work involved. TAs are good, there needs to be better connection between lecture material and projects, you will write a lot and the participation portion is not that bad.

I took this over the summer and paired it with AIES thinking a medium and an easy class would go well together. Overall, I was correct. Lots of people have outlined the course contents and what you need to do. For a point of reference, I finished with a 93% for the course. Here are some thoughts on the course:

1. You will write a lot of reports. I found them to be better than I expected. You are given a rubric to follow. If you follow the rubric, format the report so that sections match the rubric sections and you answer the questions in the rubric in each section where they are asked (even if you occasionally repeat yourself) you will be fine. After the first one or two, you get the hang of what they want. I can't prove it but adding in diagrams and flowcharts seem to help the grade.

2. There is a lot of (busy)work. Canvas shows 31 graded assignments ranging from a few course surveys to the final exam and final project. You will be busy with multiple things due each week.

3. There is a lot of "noise" in the course. Ed is filled with lots of extra posts on topical things that are not required but meant to build an online community and discuss ideas. You can decide if they are worth participating in.

4. The lectures are somewhat dated and Dr.Joyner is working on updates. He released a few of them this semester for us to preview and comment on.

5. The lectures are very much at the 30,000 ft. level. They cover a topic with examples using Ed Lessons. When you are done, you know how they work but there is no practical "how would I implement this?" or "What does the algorithm for that look like?"

6. The final project - ARC-AGI was a lot of fun. Unfortunately, many students (including myself) use a more brute force approach to solving the project because there isn't any guidance on how we might connect the dots from the 30,000 ft. level to the 10 ft. level of implementation. On the project implementation I got a 73%. Because the project implementation weight is only 7.5%, it wasn't worth the remaining hours I had left to get to 100%. Time was better spent on the report (which I got a 100% on) which isn't directly linked to how well you did the implementation but is heavily linked to how well you can explain why you chose the method you did, where it worked well, where it struggled, etc. I suggested they cut the project form 96 test cases to 60 for the summer due to the compressed schedule.

7. Participation - You need 90 points to get full participation credit. If you participate on the forums, ask questions, answer questions, make comments, you can get 40 points. If you join one of the reading groups and make it to all five meetings, you can get 25 points. They have a wide range of books and I found this to be a highlight of the course. That leaves you with 25 points to get. Each peer review that you leave if done in a timely manner is 1.5 points and they take about ten minutes to write. My suggestion is to do reviews on the ARC-AGI reports that you write for each of the five deliverables. You can do extra ones up front and get done early. This is a good way to see how others are implementing their project and get ideas you can use for yours. You can get the last 25 points from doing 17 or 18 peer reviews. I found them to be helpful.

8. With so many small assignments, one bad assignment is not going to kill your grade. I see this as a positive.

This course is an absolute nice choice for those choosing the AI track (formerly the II track). After taking it, I found it to be significantly easier than the ML and AI courses, and getting an A is quite achievable. I spent about 4–8 hours per week on this course. Only one or two of the mini-projects were relatively challenging. I recommend that those who don't want to spend too much time directly take a case-by-case approach for the AGI project, rather than trying to develop a strong generalized AI method (this is for efficiency — if you’re passionate about KBAI, then you should go for it).

As for the exam, since it allows the use of any online resources, I barely prepared and completed it using AI (which is permitted). The exam score generally doesn’t have a decisive impact; as long as you perform well on other assignments, a score of 70 on the exam is acceptable (in my opinion, this is the lower bound for using AI without your own thinking). Similarly, it’s hard to rely on the exam alone to pull your grade back up to an A.

This is my first OMSCS course (planning on a double concentration in ML and II). My background is a software Engineer for almost 2 years, did Undergrad Major: Management information Systems, minor: computer science. Only class I took this semester.

For an intro-course to grad school and getting my brain in learning shape, this was a great course for that. Wasn't exactly easy, but not too rigorous that I would be destroyed. Initially was spending about 5 ours more than the average I selected, as I took my time to be extra careful and wanted an A to give me that confidence boost, starting Grad school. Spent less time around the middle of the semester as i had gotten acclimated to it and some assignments were objectively easier. In terms of usefulness, not super confident what the real world applications of some of the lessons learned, but does introduce you to thinking patterns, libraries, etc that i think will be super helpful in more advanced courses down the line.

I thought this course was alright. My impression was content isn't very hard to understand, so they overload you with busy work to make us feel like the course is difficult. Some of the assignments are helpful, and some are just extra. Grading at times seemed arbitrary with insufficient and/or repetitive feedback given by TAs, despite the fact that rubrics are public. That being said, I don't know if this turned out to be that difficult of a course (though Exam 2 and final paper grades are outstanding, so maybe it's too early to say) all things considered. In addition, for what it's worth, Joyner did seem gung ho about getting feedback, but it's definitely too early to say that's a good sign, as instructors gathering feedback and instructors acting on it are two very different things.