v1 used a very limited (albeit very easy and already quite impressive) form of transfer learning, e.g. take a pretrained network's 1000dim vector outputs given a bunch of images belonging to three sets (since the original was trained on Imagenet), and then just use K-NN to predict what a set "new" image falls into.
v2 does actually finetune weights of a pretrained network. At the time, it was a nice showcase how fast fast JS ML libraries were evolving.
I've done my share of research on MediaPipe[1], but had never heard of Teachable Machine. I'm curious if these efforts are related, as these products looks like they were almost intended to be used together.
I am definitely excited to see that Google is investing into more "ML at the edge" use cases, especially in the browser. If you've never heard of MediaPipe before, but this caught your eye, definitely check it out. It has seen large uptake in the VTubing community especially as it has a very performant implementation of body + face + hand pose tracking driven by BlazePose.
The teaching part is what matters, it’s training (tuning in this case) a model, not just using a model already trained for inference (which is what I assume you mean). You’re providing new data that is used to update the model. Inference across an existing multimodal model doesn’t change how it classifies in any way.
I think this is more like fine tuning an existing model to recognize features you specifically intend it to, and be light enough to run locally in a browser.
It's not even fine tuning, it's creating a model from scratch. This isn't like our modern huge models either, these tiny single-purpose models have been around for ages and are quite versatile. They're so small you can't just easily run them in the browser, but train them effectively, which is what this project lets you play around with! Super cool stuff.
Like a number of Google projects, this one lives on without any clear direction. It probably will get axed some day, but the technology in Teachable Machine today is so “old school” already that I don’t think it would be that hard for someone to recreate or improve upon.
I considered the older version but it's very limited:
> The original Teachable Machine only let you train 3 classes, whereas now you can add as many classes as you like.
I'm curious to see how far this scales, for example can I have a few hundred thousand classes? If so, what are the consequences, if any?