Back to index
OtherOne 15 week semester, around 100 hours of material·Free to follow through the public course pages

CMU 10-601 / 10-301: Introduction to Machine Learning

4.3

A more rigorous and more theory forward alternative to the usual introductory machine learning courses. The public materials are extensive enough to self study properly, which is rare for a course of this depth.

What We Liked

  • Genuine theoretical grounding including PAC learning, inductive bias and generalisation bounds
  • Programming assignments require implementing algorithms rather than calling library functions
  • Slides, notes and past exams from many semesters are all publicly posted
  • Gormley's slides are unusually self contained and readable without the lecture
  • Covers the classical toolkit properly, which most modern courses now rush past

What Could Be Better

  • Lecture videos sit behind CMU's internal system, so you are working from slides
  • The theory emphasis will frustrate people who want to build things quickly
  • Notation is dense and the mathematical demand is higher than most introductory courses
  • No coverage of modern deep learning or language models, by design
  • Autograders are not available, so the assignments need self assessment

Detailed review

There is a version of machine learning education that consists of importing scikit-learn, fitting a model, reading an accuracy score and moving on. CMU's 10-601 is the opposite of that, and it is one of the few genuinely rigorous introductory courses whose materials are open enough to follow from outside. Matt Gormley has been teaching this for years and has posted the course pages for essentially every semester. Slides, lecture notes, recitation material, homework handouts and past exams are all sitting there publicly.

That accumulated archive is the reason this course works for self study despite the lecture videos being locked to CMU's internal system. The slides are unusually complete. Gormley writes them to stand alone rather than as a prompt for a talk, so working through them is closer to reading a well structured textbook than to guessing what was said. The content covers the classical machine learning canon done properly: decision trees, k nearest neighbours, perceptron, linear and logistic regression, neural networks and backpropagation, support vector machines, kernels, learning theory, graphical models, hidden Markov models, expectation maximisation, dimensionality reduction and reinforcement learning basics.

What separates it from the popular alternatives is the theory. This course spends real time on inductive bias, on the PAC learning framework, on generalisation bounds, on the bias variance decomposition treated as mathematics rather than as an analogy about dartboards. It asks why a learning algorithm should be expected to work at all, which is a question most practical courses avoid entirely. That question turns out to matter enormously the first time a model behaves strangely in production and you need a principled hypothesis rather than a guess.

The programming assignments are implementation heavy. You write the algorithms. That is the correct pedagogy and it produces a completely different quality of understanding than the fit and predict approach. Without CMU's autograders you lose the immediate feedback, but the handouts specify expected behaviour precisely enough that you can construct your own tests, and doing so is itself a useful exercise.

The honest downsides. Video is the big one, and if you learn best by watching a lecturer, working from slides will feel like a downgrade. The mathematical density is real: this is a course where notation is used precisely and you are expected to keep up. People coming from a bootcamp background regularly find the first few weeks harder than expected.

And there is no deep learning of consequence and nothing at all about language models, which is a scope decision rather than an oversight but does mean this is a foundation rather than a destination. On sequencing, I would put this after Andrew Ng's specialisation rather than instead of it. Ng gives you intuition and momentum. Gormley gives you rigour.

Doing rigour first is possible but unnecessarily painful, and doing intuition only leaves you with the gap this course fills. My 4.3 reflects excellent material with an access friction attached. If you have already done the popular introductory courses and suspect your understanding is shallower than your confidence, this is where you go to fix that.

[ final ]

The verdict.

The best free option if you want machine learning taught as a discipline rather than as a library tutorial. Expect to work from slides rather than video.