Berkeley CS189 Spring 2025 Overview: HW1–7 with Code and Data You Can Run, Plus What Fall 2026 Looks Like

🌏 中文版

Berkeley CS189 Introduction to Machine Learning is the math-heavy ML entry point at Berkeley. It does not follow CS188 Spring 2026; the two are parallel. The Berkeley AI/ML Course Guide frames them as search/reasoning vs. statistical learning — this post makes CS189 executable.

Self-study now on Spring 2025; watch Fall 2026 for shape. Spring 2025 at people.eecs.berkeley.edu/~jrs/189s25/ keeps full notes, videos, HW1–7, code/data and past exams (A3 in this site's A0–A3 scale). Fall 2026 at eecs189.org/fa26 just published a 27-lecture calendar (Lec01 Introduction + ML Problem Framing to Lec27 Closing), but most decks and assignments are still TBD and the rotating eecs189.org domain 302s to the current term, so old URLs can 404.

How public is it

Spring 2025 earns A3 because the practice loop is closed: problem set + runnable code/data + past exams for self-checking. Fall 2026 is useful to see the arc — from the Fall 2026 Schedule: Data Tools / K-Means / KNN → Density Estimation / GMM → Linear Regression / Bias-Variance → Logistic Regression → Gradient Descent → Neural Networks → CNN / Transformers / LLM → Attention / MDP / RL → Post-training / Diffusion → Closing — consistent with the canonical ML spine, but not a releasable self-study bundle yet.

Prerequisites

Official: multivariable calculus, linear algebra, and CS70 (or consent). An off-campus checklist:

1. Linear algebra: matrix products, eigenvalues, SVD geometry; read the normal equations for least squares and ridge.
2. Probability: conditional probability, expectation, MLE/MAP, bias-variance.
3. Implementation: Python + NumPy for vectorized code and gradient checks. If shaky, shore up with CS61B Fall 2025 habits for data structures and testing.

CS189 does not list CS61B by number, but HW code assumes reproducible experiments, train/validation splits, and learning curves. Missing that hurts more than a missing course code.

HW1–7 route (Spring 2025)

HW1–7 each have a PDF and paired code/data. Work in order: notes → video → discussion → HW → past exam:

1. Data and distance: K-Means, KNN, tooling; feature–distance–decision intuition.
2. Probability and density: density estimation, GMMs, EM.
3. Linear models: least squares, regularization, bias-variance, model selection.
4. Classification: logistic regression, LDA/QDA, SVM contrasts.
5. Optimization: gradient descent convergence and step sizes incl. stochastic/mini-batch.
6. Non-linear & kernels: kernel methods, feature maps, dual views of regularization.
7. Deep models: neural nets, CNN/Transformer/LLM and generative wrap-up (mirrors Fall 2026 Lec12–Lec27).

Each HW keeps the problem–code–data triangle runnable locally; what is missing is Berkeley's hidden Gradescope tests and staff feedback — substitute with past exams and your own validation curves.

Why not chase Fall 2026 directly

eecs189.org is a rotating site already 302ing to /fa26; prior term URLs can break after the switch (the Berkeley guide documents 404s). Fall 2026 currently shows a full calendar with many TBD decks/HWs. Use Spring 2025 as the spine and cross-check Fall 2026's 27-lecture order; when a new unit appears (e.g., Post-training: Fine-tuning / LoRA / PEFT / Distillation, Diffusion), backfill from the current term.

Tonight's starter

1. Open CS189 Spring 2025 HW1, don't code — label each problem as derivation, data analysis, or programming.
2. Cross-check Fall 2026 Lec01–Lec03 (Introduction / Data Tools / Math Refresher) to patch linear algebra or probability gaps.
3. If the NumPy part of HW1 runs within ten minutes, move to HW2; otherwise patch fundamentals via CS70 notes or CS61B before jumping ahead.

References

• Berkeley CS189 Spring 2025 (~jrs, canonical A3)
• Berkeley CS189 Fall 2026 Schedule (eecs189.org/fa26)
• Berkeley AI/ML Course Guide: From CS61A to CS288
• Berkeley CS188 Spring 2026 Overview
• Global AI/CS Course Map: A0–A3 Levels
• CS70 Fall 2025
• CS61B Fall 2025
• CSDIY — Berkeley CS189