AII 600: Foundations and Practice of Machine Learning for Artificial Intelligence
MS in AI. Fall 2026. George Mason University
Textbook. Tutor setup. Submit work on Canvas.
| Instructor | Vadim Sokolov, vsokolov@gmu.edu |
| Office / hours | Nguyen Engineering Building 2242; by appointment |
| Meets | Wednesdays 7:20–10:00 pm, Enterprise Hall 276 |
| Credits | 3 |
| Recommended preparation | Calculus, probability, statistics, and some programming. |
| Enrollment | College of Engineering and Computing. Class standing Advanced to Candidacy, Graduate, Junior Plus, Non-Degree, or Senior Plus. Non-degree undergraduates may not enroll. |
Schedule
No class 25 November (Thanksgiving). Week 15 is the exam block, 7:30 to 10:15 pm. Weeks 1 through 7 are lecture and labs. Week 8 is the midterm. Weeks 9 through 14 are studio and the project. Week 15 is the defense.
| Wk | Date | Topic | Work |
|---|---|---|---|
| 1 | 26 Aug | Probability; AI/ML landscape | Tutor; Lab 1 assigned |
| 2 | 2 Sep | Distributions, Bayes, correlation | Lab 1 due |
| 3 | 9 Sep | CDFs, quantiles, method of moments | Lab 2 |
| 4 | 16 Sep | MLE, likelihood, CLT | Lab 3 |
| 5 | 23 Sep | OLS, residuals | Lab 4 |
| 6 | 30 Sep | Transforms; logistic regression; classification | Lab 5 |
| 7 | 7 Oct | Testing, predictive values, expected value | Lab 6 |
| 8 | 14 Oct | Midterm | Lab 7; Midterm (2.5 hr) |
| 9 | 21 Oct | Utility, paradoxes, Kelly | Teams lock; studio; defense practice |
| 10 | 28 Oct | Bias-variance, cross-validation | Studio; proposal |
| 11 | 4 Nov | Ridge, lasso | Studio; mid-build |
| 12 | 11 Nov | Networks, ReLU | Studio |
| 13 | 18 Nov | PPR, KANs, trees vs nets | Studio |
| 14 | 2 Dec | Fairness, auditing, the decision log | Studio; decision log |
| 15 | 9 Dec | Defenses | Defense |
University calendar: first day of classes 24 August; Labor Day 7 September; Fall Break 12 October; Thanksgiving recess 25–29 November; last day of classes 7 December; reading day 8 December; exam period 9–16 December.
What this course is
The foundations of statistics and machine learning: probability, estimation, regression, classification, decision-making under uncertainty, model selection, and deep learning.
The term is in two parts. Weeks 1 through 7 are lectures and self-check labs. Week 8 is the in-person midterm with no AI assistant. From week 9 the class flips: studio in the room, and the rest of your time is the project. The project is a real decision, on data whose provenance you have checked, with a stated loss. You choose and validate a model, compare an alternative, and defend the recommendation live in week 15. There is no report and no separate presentation.
The foundation phase is 50%: labs 10 and midterm 40. The project phase is 50%. Across the course that is 50% Understanding, 40% Judgment, and 10% Preparation. Of the 40-point midterm, 25 are Understanding and 15 are Judgment. Of the project phase, 25 points are Understanding (the defense) and 25 are Judgment (studio 10 and the defense 15).
Understanding means explaining mechanisms, interpreting results in context, applying the foundations without an assistant, predicting what changes when an assumption or condition changes, and explaining the analysis your team produced. A correct number by itself is not enough. Judgment means stating the purpose, comparing reasonable alternatives, accounting for the consequences of error, calibrating claims to evidence, and identifying what would change the choice. Preparation is the labs: submit them on time and the 10% is yours. I do not mark them. The midterm is how I check that the work became yours.
Given that AI assistants can write the code, fit the models, and make the plots, I will not grade a take-home write-up, script, or report as a document.
Using AI in this course
You are expected to use AI assistants. Getting results from one is fine. Self-check labs are Preparation: submit them on time for the 10%. I do not mark the lab. Everything else that counts is something I can still attach to you or your team: studio in weeks 9 through 14, the midterm, and the project defense.
The rules are short:
- Use whatever tools you want. There is no approved list. The midterm is the exception: no AI assistant and no computer. Pen and paper, a calculator if you want one, one double-sided cheat sheet (printed or handwritten). No laptop, phone, or watch. That exam is how I check that you can apply the basic principles and concepts on your own. Tool choice does not override data-use agreements, privacy obligations, or law. Do not upload confidential, restricted, regulated, or identifiable data to an external assistant unless the data owner has authorized that use.
- Disclose what you used and for what. No penalty attaches to the disclosure. A lab: a final cell. A studio page: one line. Each decision-log entry: one line. A team that used an assistant for everything and can defend all of it gets full marks.
- You are responsible for everything you submit, including the parts you did not write. If you cannot explain it, that is on you. You will encounter AI-produced work you do not initially understand. Interrogate it before submitting and defending it. I would rather that happen here than at work.
Course tutor
The course tutor is an assistant configured for this course. It answers R and Python questions directly, asks you a question back before explaining a concept, and will not answer “which model is better?” until the purpose of the analysis is clear. Set it up once in week 1, about ten minutes.
You can point the assistant at the URLs, without downloading anything:
Or download those two files from this page and load them in whichever assistant you use.
- Cursor, Claude Code, or a similar coding tool:
@the two URLs, or put both files in your course folder and settutor-instructions.mdas the tool’s instructions file. Keepcourse-context.mdin that folder so the tool can read it. That is the version of this with the least friction. I suggest Cursor; a $20 month ($16 if you pay for a year) is enough for the course. - Claude or ChatGPT, if your account has Projects: create a Project, paste
tutor-instructions.mdinto the project instructions, and uploadcourse-context.mdas a project file. Use that project for the term. - Assistants without Projects: start a new conversation, paste both files into the first message, and add the line “Use these as your instructions for this conversation.” Use one conversation per week and paste the files again; long threads get worse.
These files are platform-neutral, although assistants vary in how reliably they follow them. If you prefer a different one, paste the same two files into it. You can edit the instructions; it is a text file.
What you will be able to do by December
By the midterm.
- Apply and explain the basic principles of probability, estimation, regression, classification, testing, predictive values, and expected value to a new problem without an assistant, interpret the numbers in context, and say what changes if an assumption or condition changes.
- Diagnose a defective analysis from its output: name the defect, point at the evidence, and predict which way the corrected number moves.
By the end of the project.
- Choose, fit, and validate an appropriate model in R or Python, using evidence not used to fit it. Compare it with a simpler or more complex alternative, and say why that choice, including the representation, serves the stated purpose.
- State what each kind of error costs, and set evaluation measures, thresholds, and loss functions accordingly rather than by default.
- Quantify uncertainty and say what it means for the decision at hand, rather than reporting an interval and moving on.
- Establish where data came from and whether it is what it claims to be, then judge whether it supports the proposed use, who is affected, how errors fall, and who is missing from it.
- Delegate analytical work to an AI assistant and stay accountable for it: decide what to hand over, verify what comes back, and explain work you did not write.
- Defend an analytical choice under live questioning, and change your mind out loud when the questioning is right.
How a week works
Week 1. No lab due. Set up the tutor and install R or Python. In class I lecture for 20 to 30 minutes and then stop. You work a problem on your own, then compare with a neighbor. Laptop and AI are allowed. We go through it, then the next block. Four blocks a night.
Weeks 2 through 7. Before class, about 90 minutes: the self-check lab from the previous lecture. The lab is a short notebook with three or four exercises and one written question. Ask the tutor anything you got stuck on. In class the same lecture-then-problem rhythm. There is no studio and no project yet.
Week 8. Lab 7 is due before the midterm. Class is the midterm: 2.5 hours, pen and paper.
Weeks 9 through 14. Outside class you work on the project. In class, studio. Teams lock at the start of week 9, before studio that night. Week 9 is studio and ungraded defense practice; there is no assigned critique that night. Critiques run weeks 10 through 14.
- Studio. I put up a problem on that week’s topic. Your team works it with your assistants. I do not walk you through it. I move between teams. Week 11 you work on your own project.
- Critique, about half an hour most nights. Spoken, in the room. One team presents. Another team critiques it using the same five questions on the rubric: what they did, what the number means, why it supports the claim, what alternative they rejected, and what would change the decision. I assign the pairings. Most of those nights two pairs, some three, so every team presents once and critiques once. This is not a peer grade. Half the 10% studio score is the weekly page; half is the two critique appearances.
Before you leave, one page: the decision; the finding; the evidence; the alternative you rejected; the strongest check that could have failed; what would change the decision. After class, add one entry to the project decision log: the decision, the alternative you rejected, the evidence, and what would make you revisit the choice. That log is due as a whole in week 14; do not reconstruct it then.
Grading
| Half | Component | Weight | Understanding | Judgment | Preparation |
|---|---|---|---|---|---|
| Midterm | Self-check labs | 10 | 0 | 0 | 10 |
| Midterm | Midterm | 40 | 25 | 15 | 0 |
| Project | Studio | 10 | 0 | 10 | 0 |
| Project | Defense | 40 | 25 | 15 | 0 |
| Total | 100 | 50 | 40 | 10 |
The same 0–3 scale is used on Understanding and Judgment.
Understanding. 3: accurate mechanism, interpretation in context, and a correct prediction if a condition changes. 2: correct explanation, incomplete or needing a prompt. 1: repeats steps or output without saying why they work. 0: wrong, contradicted by the evidence, or cannot explain the submitted work.
Judgment. 3: names the purpose, and the loss when the analysis supports a decision; compares alternatives; calibrates the claim; and responds to a challenge for a stated reason. 2: the choice is defensible, but alternatives or consequences are thin. 1: a preference without a comparison or a decision. 0: no purpose, no rationale, or keeps defending a claim the evidence contradicts.
For critique, I use the same five questions throughout. 1. What did you do? 2. What does this number mean? 3. Why does it support your claim? 4. What reasonable alternative did you reject? 5. What evidence would make you change your decision? The first three are mostly understanding. The last two are mostly judgment. A polished answer to 1 that cannot reach 3 is not understanding. A sophisticated model with no answer to 4 is not judgment.
Self-check labs. Seven notebooks, covering weeks 1 through 7. Nothing is due at the first meeting. Each lab is due before the next class; Lab 7 is due before the midterm. Credit for on-time submission: turn it in and those points are yours. There is no dropped lab and no late lab. A missed deadline is those points gone; each lab is one seventh of the 10%. Lab 1 is a page plus a Jupyter notebook you submit. From Lab 2 the assignment is the Jupyter notebook itself (Python; R kernel is allowed). Either way I do not mark the lab. The problems we stop for during lecture are separate: they are not submitted and they do not count toward the 10%.
Midterm. Week 8, the whole class, 2.5 hours, pen and paper. One double-sided cheat sheet, printed or handwritten. A calculator is fine. No AI assistant, and no laptop, phone, or watch. I will post a problem bank. It covers weeks 1 through 7. This is 40% of the course: 25 Understanding, 15 Judgment. The 0–3 scale is applied to the scored problems and then scaled to 100.
Studio. Half the 10% is the six weekly pages, equally weighted, required that night. Half is the two assigned critique appearances, presenting and critiquing equally. Both halves are team scores. The page, before you leave, in this order: the decision; the finding; the evidence; the alternative you rejected; the strongest check that could have failed; what would change the decision. It is not the project decision log.
Defense. Twelve teams of 3 or 4, locked in week 9. I will form the teams so the count is twelve and the defenses fit the exam block. Eleven minutes per team, Wednesday 9 December, 7:30 to 10:15 pm. Slots posted in week 13, two minutes between teams. About two minutes for the team to state the decision, the alternative you rejected, and what would change the recommendation. Those three are the opening; the other three studio-page parts are material for questions. Then one directed question each, with follow-ups as needed, including about work you did not personally build: an explanation or a counterfactual from the project. Assistants may help you prepare; they are not in the room. Code, plots, and analysis sit behind the talk; they are not a separate submission.
This is 40% of the course. The 25 Understanding points are individual, from that exchange. The 15 Judgment points are a shared team score, from the recommendation. The midterm is the other individual Understanding sample. If those two conflict, we have a second five-minute conversation; that is to resolve insufficient evidence, not to punish you. The common foundations stay on the midterm. The 0–3 scale is applied to the scored exchange and then scaled to 100.
The most uncomfortable part of this course is being asked “why?” and not having an answer, in a room, out loud, while people wait. That happens to everyone here, including me. Week 9 includes an ungraded practice so the first time is not the graded defense.
The proposal (week 10: a real decision, a stated loss, and a check on where the data came from), the mid-build in studio (week 11), and the decision log (due week 14) are required. They have no separate points. Late is not missing: I comment when they arrive, and I post the defense score after they are in. If they are not in when the team is scheduled to defend, we do not run that slot; we reschedule once they are, by the last day of the exam period (16 December). Skipping them is not a path to the 40 points.
If the team is not working, tell me by week 11. After that we can reduce your individual scope or reassign you. Waiting until the defense is too late. I will not ask the team to grade each other.
Scores of each component are normalized to be out of 100. Grades will be posted on Canvas. This course uses Mason’s graduate regular scale: A+, A, A-, B+, B, B-, C, F. Cut-offs: 97 (A+), 93 (A), 90 (A-), 87 (B+), 82 (B), 79 (B-), 70 (C). Below 70 is F.
Computing
You may use R or Python. Install one in week 1. The notes are mostly R; Module 6 includes some Python. Use the language you will defend. Also set up git (a course folder with notes and finalproject tracked), markdown, Jupyter, and the course tutor. Lab 1 is that setup plus the first probability; submit a Jupyter notebook on Canvas. From Lab 2 the assignment is the notebook. Start noticing datasets you have access to through work, research, or an organisation you are part of. Teams lock at the start of week 9, and the proposal is due week 10.
University policies
Mason requires every course to make available the Common Course Policies (August 2026). They cover academic standards, accommodations for students with disabilities, FERPA and the use of Mason email, and Title IX. That addendum is part of this syllabus. The AI rules in this document are course-specific and sit on top of the university academic standards; they do not replace them. I am a required reporter under the Title IX policy in that addendum. Confidential options are listed there.