CS231n 2025 Lecture Notes
Deep Learning for Computer Vision — Stanford University, Spring 2025
CS231n lecture notes · Stanford · Spring 2025 · Unofficial
Eighteen lectures
of Stanford CS231n, written up.
Vision Transformers, contrastive pretraining, latent diffusion, segmentation foundation models, vision–language models — written from the 2025 lectures, with every result linked to the paper it came from.
Each lecture is a standalone page written to be read straight through rather than skimmed. Sections are numbered and individually linkable, so you can share a link to exactly the part you mean, and equations and figures are cross-referenced so a later lecture can point back at the thing it depends on.
Lectures
The course runs to 18 lectures, and all of them are written. The roadmap groups them by theme.
Lecture 1
Why vision is the hard part of intelligence, how thirty years of geometry stalled on it, and what changed in 2012.
Complete
Lecture 2
Image Classification with Linear Classifiers
The classification setup, nearest neighbours, and why linear models are the right place to start.
Complete
Lecture 3
Regularization and Optimization
Why the data loss underdetermines a classifier, and how gradient descent and its variants find one anyway.
Complete
Lecture 4
Neural Networks and Backpropagation
What a hidden layer buys you, and how the chain rule on a graph gets the gradient of anything.
Complete
Lecture 6
VGG, ResNet, and the recipe that gets a deep network to train at all.
Complete
Lecture 7
A state carried across time, one set of weights reused at every step, and the gradient problem that reuse creates.
Complete
Lecture 8
Removing a bottleneck from a translation model, and the general-purpose primitive that fell out of it.
Complete
Lecture 9
Object Detection and Segmentation
What changes when the answer is no longer a single label, and how to ask a trained network where it was looking.
Complete
Lecture 10
What a fourth axis costs and what it buys, from clip training to 3D convolution and the methods that reach past a few seconds.
Complete
Lecture 11
Large-Scale Distributed Training
What a GPU cluster actually is, and the five ways of cutting a transformer apart so that all of it has useful work to do.
Complete
Lecture 12
Inventing tasks whose answers the data already knows, and the contrastive objective that eventually replaced all of them.
Complete
Lecture 13
Modelling the distribution over images rather than the label given one: the chain rule, and the bound that buys a latent space.
Complete
Lecture 14
Adversarial training and diffusion: a learned loss, a regression on noise, and the pipeline that ended up using both.
Complete
Lecture 15
Geometry has no canonical grid, so every method starts by choosing one. The map of those choices, and the decade-long tour across it.
Complete
Lecture 16
Text was the supervision signal vision had been missing. What that bought, what it quietly failed to learn, and the four ways of building on it.
Complete
Lecture 17
A robot’s output changes its next input, which breaks the assumption every earlier chapter rests on. Four answers, and what each one costs.
Complete
Lecture 18
What the field is for. Two claims the closing lecture states informally, made precise on the contingency table underneath both.
Complete
What these notes are
The official cs231n.github.io notes are excellent — and they stopped being updated in 2017. Everything that has since reshaped computer vision is missing from them.
These notes are written from the Spring 2025 lecture recordings and slides, not summarised from the 2017 notes. That means they follow what the lecturers actually emphasised in 2025, including the parts that changed.
Each page aims to be:
- Self-contained — you should not need the video open alongside it.
- Honest about depth — core ideas get full derivations; peripheral ones get an intuition and a link to the paper.
- Linked to sources — every result points at the original paper, not at a secondary summary.
Roadmap
Lectures are published as they are written, in course order, following the 2025 syllabus.
- Fundamentals — introduction, image classification, optimisation, backpropagation
- Convolutional networks — CNNs, architectures from AlexNet to ConvNeXt
- Sequence models — RNNs, attention, Transformers, ViT
- Vision tasks — detection, segmentation, video, 3D
- Generative models — VAEs, GANs, diffusion
- Advanced topics — NeRF, vision-language models, robot learning, ethics
Contributing
Found an error? Corrections are genuinely welcome — there is an Edit this page link at the bottom of every lecture, which opens a pull request in your browser. For anything larger, open an issue.
Disclaimer
This is an unofficial resource, not affiliated with Stanford University. For official course materials see cs231n.stanford.edu.
The notes are released under the MIT licence, which covers the writing and the diagrams made for this project. Figures reproduced from research papers and lecture slides remain the property of their original authors and are credited in their captions — see NOTICE.