RETRIEVE, THEN REVEAL
One question at a time.
All 68 cards include related lecture slide screenshots. Reveal an answer, then choose a slide or open it full-size. Original exercises and mathematical clarifications are labeled.
FROM OUR STUDY SESSIONS
Questions we've gone through.
Two questions from the current session, four answered prompts from August 26, and two reconstructed warm-ups from September 21. Open a question to check its answer and original slide evidence.
FRESH QUESTIONS · ANSWERS HIDDEN
Try it without the cards.
15 application questions followed by 5 graduate challenges in representation, Ridge, distance metrics and optimization. Write your reasoning, then compare with the key. These are practice questions, not predictions of the graduate exam.
TRACE EVERY ANSWER
Sources & study files.
What is verified
The question review includes actual images or excerpts from the lecture exports. Citations use 1-based PDF pages, HTML slide sections, or PowerPoint slide numbers. Related figures are labeled when they do not directly illustrate the question.
Recovered slides and source corrections
Lecture 02 Data Representation (20 pages), Lecture 04 Classification and KNN (35 pages), and Lecture 07 Ridge Regression (16 pages) are now available as tagged PDFs, with original PowerPoint backups. The links above open their Canvas originals. The Ridge deck has inconsistent norm notation; its numeric example and derivation establish the squared L2 penalty. See C68. C66 records the cosine example’s intermediate arithmetic error.
The 44-item instructor checklist still governs the main review. Deeper lecture practice is labeled separately; complete coverage of every possible graduate exam question is not established.
Your graduate exam
Junggab Son’s September 29 announcement: October 5, 2026 in HOS 384. Arrive at least 10 minutes early; graduate students sit in the first two rows. Graduate and undergraduate students receive different exam versions. The exact graduate questions are not published here.
Current guide and earlier downloads
The current HTML, CSV and Anki files include the recovered sources and graduate practice. The October 1 DOCX/PDF downloads are preserved snapshots and still contain the earlier source gaps.
Card progress and practice responses stay in this browser. Nothing is sent to an external service by the study app.
t-SNE, one idea at a time.
A visual supplement. Start with an answer of your own, then reveal the picture and explanation. This is extra study material, not an expansion of the instructor’s exam checklist.
Try a tiny t-SNE run: move the dots yourself
45 synthetic points, three features. The first two features overlap; the third separates three groups. t-SNE uses all three. Colors and shapes identify the groups for you; the optimizer never sees those labels.
Left: only input features 1 and 2. Right: learned map; its view auto-fits each update, so scale can change. A circle, square or triangle keeps its group identity. Nearby points matter more than absolute position or cluster spacing.
Exact demo settings and limitations
Assistant-created exact t-SNE implementation: Gaussian neighbor probabilities, symmetric P, Student-t Q, KL loss and its analytical gradient. Learning rate 25; momentum 0.5 for the first 100 steps, then 0.8; centered random initialization near zero. No early exaggeration or adaptive gains. This is a small teaching implementation, not a replacement for a production library. Runs stop at 1,000 steps. Loss need not fall at every update.
Start page: the Wikipedia article you shared. Explanations checked against the original paper, author FAQ, Distill’s interpretation experiments and scikit-learn documentation. Read October 2, 2026. Diagrams and toy calculations are original teaching examples.