Math 457: Introduction to Statistical Learning

Fall 2026 · Binghamton University

Preliminary. This syllabus is subject to change; some details (room, office hours, and a few dates) are still to be determined, and the grading weights and schedule below are proposed. Standard university statements will be appended before the final version is posted.

Instructor

Vladislav Kargin (vkargin@binghamton.edu)

Meetings

TR 3:15–4:45, in room LN 2409

Office hours

TR 2:00-3:00 in WH-136, or by arrangement over Zoom

TA

Hiten Malhotra, undergraduate teaching assistant — office hours W 1:30 -- 3:30 in room WH 227 (undergraduate lounge). Hiten took this course in Fall 2025. This is a place to work through assigned problems and homework questions with someone who has recently been through the same material; it does not introduce new material, which happens only in class. Hiten grades the computational homework; all quizzes, exams, and project work are graded by me.

This is a 4-credit course: in addition to the scheduled meetings, students are expected to do at least 9.5 hours of course-related work each week. Because the course uses a hybrid format (see below), a substantial part of that time is first-exposure reading and preparation done before class — assigned readings and course notes, short pre-class reading quizzes, computational work, studying for quizzes and exams, and preparing the project.

Prerequisite

• Scientific programming in a language such as R, Matlab, or Python.
• Linear regression and statistical inference.
• Matrix algebra, preferably including orthogonality, eigenvalues and eigenvectors, and singular value decomposition.

A short, ungraded math self-check in the first week will help you (and me) see where a refresh is needed; optional refresher material and the weekly problem session are there to close gaps.

How this course works (hybrid format)

Because our contact time is 3 hours per week, this course uses a hybrid format that combines short in-class lectures with active in-class work. Before each class you will do a short assigned reading (from ISLP, Bishop, or the course notes) and a brief low-stakes reading quiz. That preparation is what lets our class time be spent well: a typical meeting opens with a quick reading check and warm-up, moves to a focused lecture on the harder parts of the topic — the places where a live explanation beats the book — and then turns to the work that needs us in the same room: working through model choices, critiquing analyses, reading and debugging code, and short in-class quizzes on your own work.

The mix shifts with the topic. Where the textbook is thin or fragmented (transformers, for example), expect more lecture; where the reading already does the job, expect more hands-on work. What class will not do is simply re-read the textbook back to you — the reading is assumed, and the lecture builds on it rather than repeating it. Keeping up with the pre-class material is therefore essential rather than optional: it is what makes the rest of the class possible.

Description

This course is a survey of statistical learning methods for supervised and unsupervised learning, with an emphasis on judgment: when and why a method is appropriate, not only how to run it. Topics include regression with regularization (ridge, LASSO) and model selection; resampling and cross-validation; classification, logistic regression, and model evaluation (ROC/AUC, calibration); tree-based methods and ensembles (random forests, boosting); support vector machines and kernels; principal component analysis and clustering; and modern deep learning, with neural networks, convolutional networks, transfer learning, and transformers and attention treated as central. A recurring thread is the critical evaluation of analyses — including AI-generated ones — for leakage, validation, model choice, and interpretation.

Learning outcomes

Students will learn how and when to apply statistical learning techniques, their comparative strengths and weaknesses, and how to critically evaluate the performance of learning algorithms. Students completing this course should be able to:

• process and visualize different data types;
• apply statistical learning methods to build predictive models and perform exploratory analysis;
• explain the underlying mechanisms of predictive models and interpret their results in context;
• properly tune, select, and validate models, and recognize leakage, class imbalance, and other failure modes;
• read, evaluate, and critique an analysis or a piece of code — including one produced by an AI tool — and identify where it is wrong;
• understand modern architectures (including transformers) at a conceptual level and judge where they are appropriate;
• work both independently and in a team to solve problems; and
• present and communicate findings effectively, in writing and in an oral defense.

Texts and resources

Text. James, Witten, Hastie and Tibshirani, 2021. An Introduction to Statistical Learning with Applications in Python (ISLP). Home page: statlearning.com (free PDF available).

Supplementary text. Bishop & Bishop, 2024. Deep Learning: Foundations and Concepts. Home page: bishopbook.com (digital version available).

Course notes. For several topics — transformers above all, where the textbook treatment is fragmented — I will provide concise course notes that serve as the primary first-exposure material. These are listed on the schedule and are required reading where assigned.

Online resources. There is an online course taught by the ISLP authors, available on edX and YouTube. It overlaps with our course but is not identical: different ordering, and some of our topics are not covered.

Software. We will use Python and Google Colab.

Communication. We will use Piazza for questions and discussion, and all class announcements will be posted there — so sign up in the first week and make sure you receive its notifications. Sign-up link: Fall 2026 Piazza link TBD. Brightspace will be used minimally, for grade recording.

AI-use policy

This course treats AI tools (ChatGPT, Claude, Copilot, and the like) as instruments you must learn to use and to judge — not something to hide. The policy is tiered by the type of work:

• Pre-class reading, low-stakes homework, and computational work: AI use is permitted and often expected. You must be able to explain and defend anything you submit — you will be asked to, in class. Add a one-line note disclosing what tool you used and for what.
• Critique work in class: you will frequently be handed an AI-generated analysis (sometimes deliberately flawed) to evaluate, debug, and correct. Here AI is the object of study, not a shortcut — and the work is done in the room, without tools.
• In-class quizzes, the midterm, the final exam, and project oral defenses: no AI, no notes, no outside assistance. This is where most of your grade comes from, precisely because it is where your own judgment is visible.

The principle: AI can produce analysis; this course assesses whether you can tell good analysis from bad. Using AI without understanding is the one approach that will not serve you here, because the graded moments are AI-free.

Assessment components

• Reading quizzes (pre-class): short, low-stakes, and frequent; the forcing function that keeps the format working. Also serve as an engagement and background signal.
• In-class quizzes: graded, closed-book, often tied to your own submitted work. The two lowest or missing quiz scores are dropped.
• Critique work (in class, not separately graded): a recurring activity in which you evaluate and repair flawed or AI-generated analyses, usually in small groups. It is not submitted for a grade on its own — but the judgment it builds is exactly what the quizzes, the final exam, and the project defense test, and quiz questions are frequently drawn from cases we have worked through together.
• Homework: low-stakes and formative — done to learn, discussed with classmates, but written up individually. Computational assignments via Google Colab / Google Classroom (class code TBD); some conceptual work via Gradescope. A short DataCamp component may be assigned for Python onboarding (waivable with proof of prior Python coursework).
• Midterm and final exam: in-class / proctored. The midterm covers the classical core (through logistic regression). The final is weighted toward the second half of the course (deep learning, transformers, trees, unsupervised methods, SVM) and is built around judgment and critique — evaluating analyses, choosing and justifying methods, and finding flaws — rather than mechanical recall.
• Project: see below.

Project

A group project (groups of 3–4) using real data. Deliverables: a proposal, a preliminary report, slides, a final report, and an oral defense in which each member explains their own results and decisions. Negative or inconclusive results are fine if the question is real and the analysis is honest; originality of the question and the quality of the data story matter as much as method correctness.

Grading (proposed weights)

• In-class & reading quizzes — 20%
• Homework (formative) — 5%
• Midterm exam — 20%
• Project (reports, slides, oral defense) — 30%
• Final exam — 25%

Letter grades will be assigned on a scale set after the course ends, but you are guaranteed at least: A for ≥ 90, A− for ≥ 85, B+ for ≥ 80, B for ≥ 75, B− for ≥ 70, C+ for ≥ 65, C for ≥ 60, C− for ≥ 55, D for ≥ 50.

Tentative schedule (key dates)

A detailed week-by-week topic schedule is distributed separately.

Standard university statements on accessibility/disability accommodations, academic honesty, and religious observance will be appended before posting.