Training your own image-recognition AI model with Teachable Machine

The goal of the exercise

Teachable Machine is Google's no-code, browser-based tool for building machine-learning models. In this exercise you do not instruct a finished AI: you assemble the training examples, you train the model, and you measure what it actually learned. The fictional SwiftBox Logistics Ltd. wants an experimental pre-sorting system that places an object held up to the camera into one of three categories: Box, Envelope and Other. You build two versions of the model: a deliberately weak V1 in which every category sits on a different background, and an improved V2 trained on more varied data. The goal is not perfect recognition but the experience that an AI learns from what we show it, and that a high percentage does not amount to a reliable truth.

In this exercise you learn: how to define unambiguous categories for a classification problem; how to collect training data, and why it distorts the result when the background always travels with the category; how to train a model in the Teachable Machine interface; how to test with examples the model has never seen; how to calculate a hit rate and build a confusion matrix; how to diagnose the likely cause of a mistake; how to improve a model with targeted extra data; how to read a confidence value correctly; and how to formulate a human-review rule for an uncertain AI decision.

What will you need?

• A computer or laptop with a stable internet connection
• A modern desktop browser, preferably Google Chrome