Back to index
OtherSelf-paced, around 16 lectures plus assignments·Free (materials and lectures)

Stanford CS231n: Deep Learning for Computer Vision

4.8

The course that taught a generation of researchers how convolutional networks actually work. The lecture notes alone are worth more than most paid courses, and the assignments make you build the maths by hand before you touch a framework.

What We Liked

  • The published lecture notes are some of the clearest writing in the field
  • Assignments make you implement backprop and conv layers from scratch in NumPy
  • Takes you from a softmax classifier all the way to attention and transformers
  • Everything that matters is free, the notes, the slides, and the older lecture videos

What Could Be Better

  • You need solid Python and a real grasp of calculus and linear algebra first
  • The freely posted video sets lag the live course, so some newer material is notes only
  • No grading or certificate unless you are an enrolled Stanford student

Detailed review

CS231n has a reputation, and for once the reputation is earned. It started life as Fei-Fei Li, Andrej Karpathy, and Justin Johnson's computer vision course and it became the thing half the field quietly learned from, usually through the lecture notes posted at cs231n.github.io. Those notes are the reason to come here. They explain backpropagation, the bias variance trade off, convolutional layers, and optimisation with a clarity that most textbooks never reach, and they do it without hiding behind notation.

The arc of the course is the real value. You start with image classification framed as a simple linear classifier, you suffer through implementing a softmax and an SVM loss by hand, then you build a two layer network and write the backward pass yourself in NumPy before any framework shows up to do it for you. That order matters, because by the time you are allowed to use PyTorch you actually know what it is doing under the hood, which is exactly the understanding that separates people who can debug a model from people who can only restart the training run and hope. From there it moves through convolutional architectures, training tricks that actually matter in practice, recurrent networks, and on into attention and transformer based vision, so it has kept pace with where the field went rather than staying frozen in the ImageNet era.

The honest caveats are about access and prerequisites. This is a genuine graduate level course, so if your linear algebra is shaky or you have never written a gradient by hand you will hit a wall in the first assignment. And because Stanford only posts video sets periodically, the freely available lectures usually trail the current syllabus, so for the newest topics you are reading slides and notes rather than watching a polished talk. None of that changes my view.

Work through the assignments honestly, resist the urge to copy solutions, and you come out the other side understanding deep learning in a way that no amount of API tutorials will give you.

[ final ]

The verdict.

If you want to genuinely understand deep learning for vision rather than just call a pretrained model, this is still the reference. Do the assignments properly, do not just read along, because the from scratch coding is where it clicks.