Read the official Google blog post.
Some more coverage here:
AI Experiments is live.
There’s a bunch of amazing experiments on the site; but this one below is the one I spend the most time with during its early development phase.
Honestly; I never felt more out of my depth on a project than at the beginning of this one. Sat in the kickoffs with Alex, Kyle and Yotam who were deep in the weeds talking about t-SNE, dimensionality reduction, hi-dimensional space, convolutional neural networks, and supervised vs un-supervised learning. Was a full-on nose-bleed, crash course, in ML. But so worth it. Do not fear this stuff. It’s a different world to start; but after a few weeks it starts to take. So please enjoy….
The Infinite Drum Machine
Sounds are complex and vary widely. This experiment uses machine learning to organize thousands of everyday sounds. The computer wasn’t given any descriptions or tags – only the audio. Using a technique called t-SNE, the computer placed similar sounds closer together. You can use the map to explore neighborhoods of similar sounds and even make beats using the drum sequencer.
Here’s the explainer video:
For an extra sneak peak into the development process; here’s a video showing an earlier prototype. This one has around ~40k short samples from Freesound! For the final version we licensed ~17k.
This is one of the last projects I started working on in New York, so it’s great to see it out in the real world. Mad props to Alex, Catherine, Manny, Yotam, Eric, Jonas, Kyle, Gene and bunch of other very smart people.
And yea…. what Kyle said.
AI Experiments website:
Tweetfuel from Stinkdigital on Vimeo.
TweetFuel is an Arduino-controlled experiment that uses the Nike+ FuelBand to measure the health of our Twitter account.
This is something I meant to write about a long time ago. I also wanted to write a detailed and smart post about how and why the project came about, what it’s meant to say about Stinkdigital, how we had to get Nike onboard because we do a lot of work for them, maybe share some details about how and why we simplified the site design to make it feel less like a ‘product’ launch, how we used HTML5/JS, Arduino, Python, and the coverage we got in FWA, FastCo Create, FastCo Design, and lots of other magazines and blogs.
But time has passed, it’s was launched almost 3 months ago now…. so all you really need to know is that we got it from concept to working prototype in 5 days.
Whenever someone follows, RTs or mentions @stinkdigital, our custom made motorized kit spins a mounted Nike+ FuelBand.
So far the video has had 24,100 plays — and I’m still super happy with it as a piece of work.
Retweets count. (So do mentions.)
An experimental film in tribute to Ridley Scott’s film Blade Runner, made as a unique picture with a resolution of 60.000 x 60.000 pixels (3.6 gigapixels) using 167,819 frames from the 1982 classic.
BLADE RUNNER revisited >3.6 gigapixels from françois vautier on Vimeo.