In case you’ve been worried about the imminent takeover of self-aware robots who calculate our inevitable extinction, there’s some good news: At this point, they can be foiled by a Grateful Dead sticker.
At the most recent Neural Information Processing Systems conference in Long Beach, California, a group of Google engineers revealed that the image recognition technology in advanced machine learning systems can be completely disrupted with a simple psychedelic sticker.
One of the images the team used as an example was a banana, which the machine could easily recognize. However, when the image of a psychedelic swirl was placed next to the banana, the machine was lured away from the banana and registered that the picture was “of” the swirl.
Just as the human eye does, AI uses cognitive shortcuts to comprehend images. The eye will attribute more importance to certain pixels over others. When the eye sees a picture of a house or a boat, for example, it will know that the picture is not “of” the sky, or a cloud, or the ocean, etc., despite those images also being present. The Google engineers exploited these cognitive shortcuts by creating psychedelic images that were so mesmerizing, the AI involuntarily fixated on them. This is done on a system-specific, not image-specific basis, so the resultant scrambler patch will generally work no matter what the image recognition system is looking at.
The bad news is this may throw a wrench into the predictions of machine learning making us more secure, particularly when it comes to contraband screening at airports and other high-risk locations. But the good news is that Skynet isn’t ready for rollout just yet.
For a deeper dive into the emergence of AI in an enterprise setting and the potential impact of AI-related technologies, check out this research report by Aberdeen’s Michael Lock.