How can we make it easier to experiment with machine learning and neural networks? This experiment was a collaborative effort by friends from Støj, Creative Lab and PAIR teams at Google to test out neural network learning, an open-source library that allows web developers to train and run machine learning models locally in their browser.
The Google Creative Labs team came to us with the vision for what’s called a “Teachable Machine.” This program lets anyone explore how machine learning works in a fun, hands-on way.
We helped develop the neural network locally on any device, without sending any images to a server, with an exceptionally quick response. Because it is an experiment, there are mostly playful results like making your hand say ‘moo’, or trigger cool sounds or GIF’s by wiggling your fingers. The more people use Teachable Machine, the better results will become, ultimately challenging the creative imagination. We encountered many challenges, iterating through them as a cohesive unit. For instance, how do you capture 30 images per class, in a variety of angles or variations on your subject matter?
We brought machine learning to the people, teaching them how to embrace the process of trial and error and learn from the examples you provide to the code. This experiment lets anyone use machine learning through simple actions instead of code. The tool makes neural networks more approachable and relatively easy to try. The image recognition is powered by a neural network, made possible by Nikhil Thorat, a member of the team behind deeplearn.js. The code for this experiment is open-sourced on Github.