CS231n: Deep Learning for Computer Vision

What CS231n is

CS231n: Deep Learning for Computer Vision is Stanford's long-running course on neural networks for images. The most recent offering listed on the course site is Spring 2026. Enrolled students get the current lectures through Stanford's own systems, but the course notes are public at cs231n.github.io and recordings from earlier years are on YouTube, which is how most people outside Stanford take it.

What it covers

The public notes are organised into modules. Module 0 prepares you with a Python and NumPy tutorial that works in Jupyter or Google Colab. Module 1 covers image classification, linear classifiers, optimisation, backpropagation and how to set up and train neural networks. Module 2 covers convolutional neural networks, how to visualise what they learn and transfer learning. The lectures go further, into detection, segmentation, generative models and transformers for vision.

Is it worth your time?

Yes, if you want to understand computer vision rather than only run it. The assignments make you implement a k-nearest neighbour classifier, softmax, fully connected networks, batch normalisation, dropout and convolutional layers yourself before you rely on a framework. After that, error messages and strange loss curves in PyTorch stop being mysterious. The notes are also some of the clearest writing on CNNs anywhere, and people return to them for years.

Prerequisites

The course site asks for solid Python with NumPy, college calculus and linear algebra, and basic probability and statistics. If matrix multiplication or the chain rule are shaky, work through a maths refresher first; the course moves quickly.

How long it takes

At Stanford it runs for one term of about ten weeks. Self-paced learners usually take longer, because the assignments are substantial. Plan on a few hours per lecture plus the assignments, and do not skip them: they are where the learning happens.

Free and paid

The notes, assignments and past lecture videos are free. Credit, grading and the current term's recordings are only for enrolled Stanford students.

What you can do after it

You can explain and implement the building blocks of modern vision models, read vision papers with confidence, fine-tune pretrained CNNs, and debug training runs by reasoning about shapes, gradients and data rather than guessing.

Common questions

Is it too old? The fundamentals it teaches, backpropagation, CNNs and transfer learning, are exactly what current models are built on, and the course is refreshed each year.

Best for: programmers with some machine learning who want a rigorous, free path into computer vision.