Member-only story
Facebook Creates ‘Radioactive Images,’ and Other A.I. News From This Week
Plus an algorithm that lets drones automatically detect protests and religious celebrations.
Researchers publish hundreds of new A.I. research papers every week on arXiv, an online repository for scientific papers. Here are a few that we found interesting, and think you might, too.
Radioactive data: tracing through training
Facebook researchers created a way to generate an invisible “radioactive” mark on images in a dataset, which gives future researchers a way to trace which data was used to train a specific model. That could be useful when certain companies *cough* Clearview AI *cough* claim to have a facial recognition database scraped from images on the web. More from Facebook here.
Accelerating Object Detection by Erasing Background Activations
Deep learning takes a lot of computing power, especially when working with images. Intel researchers built an objectness mask generation (OMG) network to help predict where an object is likely to be, so compute power can be focused on a specific region of pixels rather than the whole image. And that is a LOL (Latency Optimizing Labor).
