Back to index
OtherOne semester, 22 lectures plus 6 assignments and a mini project·Free

EECS 498-007 / 598-005: Deep Learning for Computer Vision

4.8

The best free deep learning course I know of, and I do not say that lightly. Johnson explains difficult material with unusual clarity, and the assignments make you build backpropagation, convolutions and attention yourself, which is why people who finish it actually understand this stuff.

What We Liked

  • Explanation quality is exceptional, with the backpropagation and convolution lectures being the clearest public treatments of either topic
  • Assignments are implemented from scratch before you are allowed to use the convenient abstractions, which is the correct order
  • Complete lecture recordings are on YouTube, so the whole course is genuinely followable from outside
  • Covers the full arc from linear classifiers to attention, detection, segmentation, generative models and 3D vision
  • Johnson taught CS231n before this, and the pedagogy shows the benefit of a second pass at the same material

What Could Be Better

  • Recordings are from an earlier offering, so recent developments in vision models are absent
  • Assumes Python fluency, calculus and linear algebra, and the pace assumes you are not learning those alongside
  • The assignments are long, and doing them honestly takes considerably more time than watching the lectures
  • Transformer and multimodal coverage is light by current standards, since the course predates the shift
  • No grading, feedback or certificate for outside learners

Detailed review

I have watched a lot of deep learning lectures and this course sits at the top. Justin Johnson taught CS231n at Stanford before building this at Michigan, and the second attempt is better, which is what you would hope for and rarely get. The lecture on backpropagation is the one I send people to when they say they understand the chain rule but cannot explain what a computational graph is doing. The convolution lectures build up from filters to architectures without ever waving a hand over the arithmetic.

This is the difference between a course that teaches you the words and one that teaches you the mechanism. The sequence runs from image classification and linear classifiers through optimisation, neural networks, backpropagation, convolutional architectures, training practicalities, then out into detection, segmentation, recurrent models, attention, generative models, self supervised learning, 3D vision and video. The assignments are the reason people who take this course come out different. You implement a k nearest neighbour classifier, then linear classifiers, then a two layer network with your own backward pass, then convolutional networks, then work up to attention.

Only after you have built the pieces yourself are you handed the frameworks that make them easy. That ordering is the single most important pedagogical decision in the course, because a student who has debugged their own gradient computation has intuitions that a student who has only called a library never develops. The one substantial weakness is age. The public recordings come from an earlier offering, and computer vision has moved since, with vision transformers, large multimodal models and modern generative approaches now central to the field and only lightly represented here.

That does not undermine the course, because the foundations it teaches are the same foundations those newer systems rest on, and someone who genuinely understands convolutional architectures, optimisation dynamics and attention will read a modern vision paper far more easily than someone who started with the papers. It does mean this should be your first course and not your only one. Pair it with recent material on vision transformers and multimodal models afterwards. Be realistic about effort.

The assignments are long and non trivial, and the honest completion time is much higher than the lecture hours suggest. The prerequisites are also firm rather than negotiable. Python fluency, calculus and linear algebra. If you are shaky on any of them the lectures will still make sense while the assignments will not.

As with any public university course you get no grading, no feedback and no credential, so your motivation has to come from wanting to understand rather than wanting a certificate. My 4.8 is the highest rating I would give any free technical course. It is clearer than most paid alternatives, its assignments are better designed than most paid alternatives, and it costs nothing. The half point comes off for currency rather than quality, and I would still start here over anything newer, because the newer material assumes exactly what this course teaches.

[ final ]

The verdict.

If you want to learn deep learning properly and you are willing to do six real assignments, start here rather than paying for anything. Follow it with something current on vision transformers and multimodal models, because the foundations here are close to perfect and the recent frontier is missing.