Google has put out a new AI experiment in the form of the "Teachable Machine," a browser-based machine learning concepts teacher. You can use the camera on any compatible device with the right browser, from your phone to your laptop, to take part in a simple machine learning demonstration. Users get to program a simple machine to recognize visual inputs and give certain outputs, with three trainable categories. The demonstration does not go into detail about how exactly machine learning works under the surface but does show users enough that most will get the basic concept by the end of their time with the experiment.
Once the experiment connects to your device's camera, you'll be greeted by three trainable classes, being green, purple, and orange. You can train any of the three classes to recognize anything you put in front of your camera, and then respond in a predefined way when it figures out which class it's looking at. Responses include GIFs, speech, and sounds, all of which are customizable. To train a class, just hold down the button for that class, and put something in front of the camera; an object, facial expression, certain action, or just about anything else you can think of. Essentially, the program will guess which class is most similar to what is in the camera's view, once it's been trained. It will tell you, in percentages, how sure it is of what it's looking at, making it possible to intermingle classes by showing something ambiguous or something that falls between two classes. Showing the camera a red drink can for the green class and a blue one for the orange class, for example, will cause both to hover around 50% and fight back and forth for control of the output if you show the camera a purple drink can.
This machine learning experiment is one of the simplest ways for users with no understanding of how AI works to get a look at the basic concepts behind it. The machine itself runs on the deeplearn.js framework which is Java-based and should work in most browsers, and be reasonably smooth and fast on most hardware. The platform is not a deep learning model by itself, but rather a framework that almost any model can be imposed upon, allowing AI-savvy users with some Java knowledge to create browser-compatible, lightweight AI projects and experiments.