Using AnyLearning to teach machine learning

Teaching applied machine learning has a setup problem. Before a student trains anything, someone has to get CUDA working on a lab machine, or hand out cloud credits, or collect consent for uploading images, or debug why one laptop resolves a package differently from the other twenty-nine. The interesting part of the lesson arrives late, if it arrives at all.

AnyLearning is a desktop application with the runtime and the pretrained weights inside the installer. Installing it is the setup.

Why local matters in a classroom specifically

• No accounts. There is nothing to sign into and no activation key, so you are not creating logins for students, and for under-18 students you are not creating them at a third party either.
• No data-processing agreement. Student photographs, coursework images and anything recorded in the lab stay on the lab machine. There is no vendor to add to your institution's data agreements, because there is no vendor.
• No per-seat cost. It is Apache-2.0 and there is no usage meter, so a class of forty costs the same as a class of one, and a student who wants to keep experimenting after the lab closes can install it at home.

What a first session looks like

A workable ninety-minute lab: students label around fifty images of two classes, train an image classification model, and read the validation accuracy. Then they look at what the model got wrong.

That last step is the lesson. Almost always the errors trace back to the labelling: an ambiguous class boundary, a class with eight examples against another with forty, images that all share a background the model learned instead of the object. Students fix the labels and train again, and the number moves. The loop from decision about data to measured consequence is short enough to fit in one sitting, which is what makes it stick.

This is difficult to teach from a slide and easy to teach from a training run that a student caused to fail.

Mapping to a course

The tutorials are ordered so they can be taken as a sequence:

• Start with the conceptual material: what machine learning is, choosing a task, and train, validate and test. The third one is the one students skip and the one that matters most, because it is where the difference between a good number and a real number lives.
• Then a task per session: image classification is the gentlest start, object detection is the most recognisable, and handpose classification reliably gets a reaction because the input is the student's own hand.
• Auto labeling is worth introducing once students have felt how long manual annotation takes. Introducing it earlier removes the experience that makes them appreciate it.

For non-CS courses, Tabular AI and Text AI cover spreadsheet and text data without asking students to write code.

Honest limits

It is a single-user desktop application. There is no class management, no per-student accounts, no grading integration and no way to see a student's project from your own machine. Every install is standalone, so distributing a starter dataset means distributing files.

If you need submission tracking or shared review, that will have to sit outside the app.

Getting started

Install it on one machine first and run through a task end to end before the lab. It is a few gigabytes because the models ship with it, which is worth knowing if thirty machines will download it over the same campus connection. Prepare a mirrored copy on a share if that is a concern.