Practical Deep Learning for Coders — fast.ai

Jeremy Howard teaches deep learning backwards on purpose. Lesson one has you training and deploying a working image classifier before you have seen a single derivative. The theory arrives later, once you have a reason to want it. Thirty years of teaching machine learning and a stint as president and chief scientist of Kaggle sit behind that decision, and it is the single most polarising thing about the course.

Part 1 is nine lessons of about ninety minutes each, covering computer vision, natural language processing, tabular data and collaborative filtering. Part 2 adds twenty-five more lessons that go down to foundations — building the pieces from scratch. The stack is PyTorch underneath, fastai on top, with Hugging Face Transformers and Gradio for the NLP and deployment work. There's a companion book, free to read online.

What it assumes

The site is direct about it: you know how to code, about a year of experience is enough, preferably Python, and you have at least followed a high school maths course. That's a lower bar than most deep learning courses set and it's honest — you genuinely do not need linear algebra to start here.

You do need patience with a non-standard library. fastai's API is opinionated and layered in ways that are wonderful when they fit your problem and opaque when they don't.

Top-down teaching, and who hates it

Some people find lesson one exhilarating. Others find it unbearable, because a model trains and they cannot say why. If you are the kind of learner who needs the foundation before the application, this course will feel like being handed a working machine with the panel welded shut. Howard argues the motivation is worth the temporary discomfort. He is right for many people and wrong for some, and there is no way to know which you are without trying.

The bigger practical issue is currency. The current Part 1 was recorded at the University of Queensland in 2022. In deep learning that is a long time — the material predates the bulk of the current generative and LLM tooling, and the Hugging Face and Gradio APIs shown on screen have moved. Concepts hold up well; specific code and library calls often do not, and you should expect to be reading current docs alongside the videos.

Who it's for — and who should skip it

Take it if you're a working developer who wants to build things with deep learning rather than publish about it. Skip it if you want mathematical rigour first — a traditional course will serve you better. Skip it too if you specifically want to work with modern LLMs, because that is not what this recording covers.

Our take

Still the best course for getting a coder productive with neural networks fast, and increasingly showing its recording date. Take it for the mental model and the top-down philosophy, and expect to patch the code against current library versions as you go.