Where CS 182 Past Exams Actually Live
Most people looking for past exams end up on a dead-end search page with a broken link to Scribd. The Berkeley CS 182 page under Sanjeev Arora has historically posted exam materials on the course website itself, but the links tend to rot each semester after the new one drops. The most reliable place I have found is the course site at cs182 Berkeley or through the class Discord where students share scanned PDFs of the actual midterm and final sheets. The exams are not your typical plug-and-chug math test. They ask you to prove things about generalization bounds, PAC-learning frameworks, VC dimension calculations, and sometimes regret analysis for bandit-style problems depending on the semester focus. I remember one exam that asked me to derive the sample complexity bound for a hypothesis class with a specific growth function, and the trick was recognizing the class wasn't what it first appeared to be. I stared at it for ten minutes before realizing the answer required me to compute the shattering threshold rather than just quote the standard formula. I wasted about twenty points on that one before catching it. If you want the actual files, check the course GitHub repo if they maintain one, or reach out to someone currently in the section. A lot of it circulates through the Discord or the Ed forum, which I would strongly recommend you join early. The staff sometimes posts practice problems there that look suspiciously like past exam questions in structure.
How to Actually Use These Exams
Don't just read the solutions. That is the most common mistake I see. People download a PDF, look at the answer key for the proof steps, nod along thinking it makes sense, then freeze when asked to produce it themselves. The first time you take the exam, do it timed. No notes. If you cannot finish it in the allotted time, that is your real baseline. The proofs tend to follow a small set of patterns. You will see the same techniques repeated: union bounds over finite classes, Chernoff or Hoeffding inequalities applied to empirical means, and the standard symmetrization argument for Rademacher complexity. Once you recognize the pattern, the exam takes less time because you are not figuring out which tool to reach for. It is more about correctly setting up the inequality than doing exotic manipulations. I used to write out full formal proofs on practice runs, but that was a waste. On the real exam, partial credit matters and graders are looking for the correct logical structure, not elegance. I started writing the key inequality steps with enough justification that a tired graders at 5 PM could follow the argument, skipping the tedious algebra that does not earn points.
Common Pitfalls
The biggest issue is that students confuse the finite-class bound with the VC-dimension bound and apply the wrong one. If the problem gives you a VC dimension, use the growth function bound. If it gives you a finite hypothesis set, use the union bound directly. Mixing those up costs easy points. Another thing: the convention for the margin in some semesters differs from the textbook. Check what the lecture notes define, not what you memorized from a different source. The exams also occasionally include a part that requires you to state whether a result still holds if you relax a condition, like dropping the independence assumption. Those are usually straightforward if you actually followed the proof in lecture. If you only memorized the final bound without understanding the step where independence was used, you will guess wrong.
Get the Full Details
What the Exams Don't Cover
They do not typically ask about deep learning architecture analysis or reinforcement learning convergence proofs. The course is theory focused, so the exams stay within computational learning theory: PAC learning, structural risk minimization, online learning regret bounds, and basic generalization theory. Don't spend hours studying topics that have never shown up. The difficulty is consistent semester to semester. The core material does not change much. What changes is which problems they feel like asking. Sometimes they lean heavily onRademacher complexity. Other times it is all about covering numbers and fat-shattering. A week before the exam, scan the homework problems and lecture examples. The exam format mirrors those closely. Download the past exams, simulate exam conditions twice before the real thing, and review the official solution writeups only after you have attempted the problem yourself. That is the sequence that works.
