If AI’s mastering anything, it’s the ability to “see,” as well as generate images. Neural networks—algorithms that harness big data—are fantastic at figuring out which IRL objects are which, for example. And they can even complete masterpiece paintings. Now, again at the intersection of silicon brains and visualizations, we have a program that can take still images and give them moving elements. Incidentally, like an acid trip would. (We’re guessing.)
YouTuber Károly Zsolnai-Fehér, who goes by Two Minute Papers, recently posted the above video to his channel. For those unfamiliar, Zsolnai-Fehér provides what amounts to a play-by-play from the cutting-edge of artificial intelligence; specifically in regards to neural networks and machine learning. (If you want to know what roaming quad-bots or Snapchat may be up to in the near future, keep track of this channel.)
In this video, Zsolnai-Fehér explains an AI which a team of researchers at The Interdisciplinary Center Herzliya in Herzliya, Israel, created. Tavi Halperin, a PhD in computer vision and deep learning, led the team, which described how the program works in a paper available to the public. Halperin’s team has even created an app that deploys the AI.
As Halperin notes in the video above, the AI essentially takes whatever portion of an image a user has selected and applies an algorithm that identifies it as individual element. The AI then loops the element to create the “constantly moving” effect. The logic behind the AI gets tricky quickly, but it boils down to sampling a “slice of pixels” in the target area, finding repetitions, then figuring out how to loop the target in a way with which Otto from
Not only is the program itself quite impressive, but it also apparently needs significantly less compute power to run. Zsolnai-Fehér notes that a previous program performing the same task required a “honkin’ big” graphics card. This program, on the other hand, only requires the one in the iPhone. Speaking of which, Halperin et al. have made the program available in the form of an app: Motionleap. It’s kind of difficult to master however, so you may have to loop through the step-by-step instructions a few times.