Teaching a Model Your Own Exercise Form

Most interactive machine learning demos hand you a record button and hope for the best. The problem is that a non-expert has no idea what makes a good training set: how many examples, how varied, how balanced across classes, or what to do when the model is confidently wrong.

This app makes that the interface's job. You teach it your own exercise poses through a webcam, and while you do, it shows a live prediction overlay so you can see what the model currently believes, gives feedback on how balanced your dataset is across the poses you have taught, and lets you correct it mid-workout when it misreads a rep. The teaching loop is the product, not a setup step before the product.

It runs entirely in the browser on TensorFlow.js and MoveNet, with everything kept in local storage: no backend, no account, and no video leaving the machine, which for a webcam application pointed at someone exercising is a design requirement rather than a nice property.

The design is grounded in eight cited papers from the interactive-ML and HCI literature, and the report proposes a user study to test whether the guided teaching loop actually produces better classifiers than an unguided one. That study was never run, so the interaction design is argued rather than demonstrated.