Which Things Can a Teachable Machine Tell Apart?
The research question
Is it easier for an AI to tell apart things that look very different, or things that look alike?
Abstract
I trained a teachable-machine model on different pairs of objects and measured its accuracy. It did much better on objects that looked very different.
Background
An AI image model learns patterns from example pictures. I wondered whether some things are harder for it to learn than others.
What I did
I trained the model on three pairs: a spoon versus a fork, a red ball versus a blue ball, and two very similar leaves. I tested each with new pictures.
What I found
It was nearly perfect on the ball colours, good on spoon versus fork, and weakest on the two similar leaves.
What's next
I want to find out whether giving the model many more leaf pictures helps it tell them apart.
Takeaway
AI learns easy differences fast, but it needs much more practice for things that look alike.