Machines are becoming increasingly adept at creating content. Whether it be news articles, poetry, or visual art, computers are learning how to mimic human creativity in novel — and sometimes disturbing — ways.
Text-based content is fairly easy for computers to generate. Anyone who has used a smartphone to text knows that operating systems are pretty savvy in predicting speech patterns. But videos and other visual mediums are a little more challenging — not only does a computer need to predict a logical thought, it also needs to visualizethat thought in a coherent manner.
It’s a challenge that came to light last week with the revelation that Youtube is home to some decidedly unsettling children’s videos. They feature popular characters like Elsa from “Frozen” or Spiderman and the kind of simple songs and colorful graphics every parent is familiar with. Watch these videos for more than a few seconds, though, and it’s hard not to feel creeped out.
Though some feature scenes of explicit violence, there’s a certain “wrongness” to most of them, as if they were alien content attempting to masquerade as “human” creations. Which, essentially, is what some of them are.
Writer James Bridle recently touched on the topic in a popular Mediumarticle. With so many kids watching YouTube videos, he explains, certain channels are pumping out auto-generated content to earn advertising dollars. Some videos seem to have benefited from human input, but others are clearly automated jumbles.
It’s about as far as you can be from the dedicated — and human — teams crafting beloved children’s movies at Disney and Pixar. It’s also the result of an emerging effort to shift some of the burden of video production to computers. It’s something that’s attracted the attention of both artists and researchers, and we’re sure to see more in the future. Whether it’s recreating a deceased “Star Wars” character or churning out children’s videos for a quick buck, the industry is still in its infancy.
One way that computers can “cheat” in creating believable visual content is by extrapolating from an already existing image or video. The combination of an existing starting point and a bit of training allows the computer to create video.
In the world of auto-generated visual content, that training usually comes from absorbing content from other videos — lotsof videos. In this study out of MIT and the University of Maryland Baltimore County, the system was trained on a year’s worth of video content.
In that case, a still image was used to generate small videos predicting what would happen next in the scene. For example, images of beaches result in crashing waves and photos of people become videos of walking or running. Due to the shaky, low-resolution quality of the video, they’re all pretty creepy (especially the babies), but the study is promising.
