CS 7641: Machine Learning
Review and Retrospective
Disclaimer
The views and opinions expressed in this post are solely my own and do not reflect those of Georgia Tech, the OMSCS program, or any affiliated instructors, TAs, or staff.
Instructor: Charles Isbell & Michael Littman
Semester: Spring 2023
Overall Rating: 6.1 / 10 👎
✅ Pros
- Covers a wide breadth of classic ML algorithms in one course
- Four large‑scale assignments force you understand the algorithms
❌ Cons
- Lectures wander into advanced tangents before establishing the fundamentals, which came acroes as self‑indulgent and disorganized; the core concepts are often lost in the digressions
- Course materials and starter libraries (e.g., ABAGAIL) are outdated or incomplete, so you spend hours hacking around them
- Assignment specs are vague, yet grading is strict
- The sheer volume, combined with disorganized and poorly presented content, turns it into an unnecessary slog
🕒 Time Commitment
Plan on 20 – 30 hours per week. Taking it alongside another moderately heavy course (I paired it with CS 7646) felt like chugging warm milk on a scorching afternoon; technically doable, but you’ll regret the decision.
📝 Grade Breakdown
| Component | Weight | Details |
|---|---|---|
| Assignments | 60 % | A1: Supervised Learning (15 %) A2: Randomized Optimization (15 %) A3: Unsupervised Learning & Dimensionality Reduction (15 %) A4: Reinforcement Learning (15 %) |
| Final Exam | 30 % | Cumulative, closed‑book, proctored via Honorlock |
| Reading/Writing Quiz | 5 % | Unlimited attempts, due Week 2 |
| Hypothesis Quiz | 5 % | Due Week 2 |
| Extra Credit | +1 – 2 % | Optional problem set and Ed participation |
✍️ Assignments
A1 Supervised Learning
Decision trees, regression, and ensemble methods; heavy experiment matrix.
A2 Randomized Optimization
Hill‑climb, simulated annealing, GA, MIMIC; lots of parameter sweeps.
A3 Unsupervised Learning
Clustering, PCA/ICA; evaluate metrics vs. dimensionality reduction.
A4 Reinforcement Learning
Implement Q‑learning on a discrete environment and analyse policy performance.
📖 Exam
Single 30% final exam covering the entire syllabus. Expect theory questions (PAC, VC‑dim, information theory) alongside algorithm mechanics. Formula sheets are not allowed.
📚 Course Content
Decision Trees
From basic classification/regression splits to ID3, information gain, handling continuous attributes, and limits on expressiveness.
Regression & Function Approximation
Linear and polynomial regression, choosing model order, error metrics, cross‑validation, and the bias–variance trade‑off.
Neural Networks
Perceptrons, sigmoid units, gradient‑descent training, XOR networks, and early multilayer‑perceptron heuristics.
Instance‑Based Learning (k‑NN)
Distance metrics, domain weighting, curse of dimensionality, and practical considerations for lazy learners.
Ensemble Methods (Boosting & Bagging)
Weak vs. strong learners, AdaBoost mechanics, bias–variance benefits, and illustrative code walk‑throughs.
Kernel Methods & Support Vector Machines
Optimal separating hyperplanes, soft margins, the kernel trick, and large‑margin intuition.
Randomized Optimization
Hill‑climbing, simulated annealing, genetic algorithms, and MIMIC for search and hyper‑parameter exploration.
Clustering & Unsupervised Learning
Single‑linkage, k‑means, Gaussian mixtures via EM, soft clustering, and performance properties.
Feature Engineering
Feature‑selection filters/wrappers, search heuristics, PCA, ICA, alternative transformations, and relevance vs. usefulness.
Information Theory
Entropy, mutual information, KL‑divergence, and coding theory basics applied to ML data representation.
Bayesian Learning & Inference
Bayes rule, Naïve Bayes, belief networks, MDL principle, sampling, and inference algorithms.
Computational Learning Theory & VC Dimension
PAC learning, mistake bounds, hypothesis‑space capacity, and sample‑complexity results.
Markov Decision Processes (MDPs)
States, actions, rewards, discounting, policy evaluation, and dynamic‑programming solutions.
Reinforcement Learning
Q‑learning, value‑function approximation, exploration vs. exploitation, and convergence guarantees.
Game Theory & Multi‑Agent RL
Minimax, mixed strategies, repeated games, folk theorems, stochastic games, and learning in multi‑agent settings.
💬 Participation & Interaction
Live office hours are indispensable, they focus on defining assignment expectations and grading details. A study group proved valuable, and while the class Slack channel is mostly chatter, it does surface the occasional gem of useful information.
💭 Final Thoughts
CS 7641 aims to be the OMSCS survey of “all things ML,” but poor organization, distracting lectures, and incomplete and aging tooling make the course challenging for the wrong reasons. Consider pairing it with a light elective, taking it by itself, or postponing it and hope for a course overhaul.