MathWorks has been quietly running one of the better free education programmes in technical software for years and hardly anyone outside engineering departments notices. The Onramps are short, free, browser-based courses that require no local install, and Deep Learning Onramp is the one aimed at neural networks. The structure is sensible and it resists the temptation most introductions give in to. Rather than starting with gradient descent and backpropagation and losing everyone before anything works, it starts by handing you a pretrained network and having you classify images with it.
You get a result in the first few minutes. Then it works backwards into how image collections are organised with datastores, how you prepare data for transfer learning, how you retrain the final layers on your own categories, and how you evaluate what came out. Transfer learning is exactly the right first real technique to teach, because it is how a large fraction of applied computer vision actually gets done. The delivery deserves specific credit.
Exercises run in a browser-hosted MATLAB with automated feedback, which means when you get something wrong you are told immediately and specifically rather than staring at an error and guessing. Anyone who has watched a beginner lose an evening to an environment problem before writing a line of model code will understand why removing setup entirely is worth more than another hour of content. The certificate is free too, which is a pleasant contrast to the vendors covered elsewhere on this site who give away the course and charge a few hundred dollars for the badge. Now the caveat that decides whether this is right for you.
It is MATLAB. Every line of syntax, every function name, every idiom is MATLAB-specific. The conceptual understanding transfers, transfer learning is transfer learning, but nothing you type does. The research world and the overwhelming majority of industry jobs run on Python and PyTorch.
If your goal is employment as an ML engineer, two hours here buys you less than two hours of fast.ai or the PyTorch tutorials. The people this is genuinely right for are those already living in MATLAB, which means a great many mechanical, electrical, aerospace, biomedical and control engineers, plus a lot of academic labs. For them this is not a detour, it is the fastest available path to using deep learning inside the tooling they already have, alongside Simulink models and the rest of their existing work. The scope is narrow, image classification and nothing else.
No sequence models, no attention, no generative anything. And the moment you want to go further you are into licensed territory, where MATLAB plus the Deep Learning Toolbox costs real money outside an academic licence. Four point zero. Free, well-built, honest about what it is, and rated on the assumption you are in the audience it was designed for.