What you need to know
- Microsoft researchers trained a deep learning model to hunt down bugs — without data.
- The model, called BugLab, involves pitting two models against each other: One that spreads hidden bugs, and another that's driven to find them.
- The two attempt to outwit each other and, in the process, train each other to be better at the joint pursuit of bug spotting.
You don't have to be deeply learned to appreciate the deep learning wizardry Microsoft's been up to. Redmond's trusty researchers are up to their usual high-tech antics and, in this case, that means discovering the secret of how to train a deep-learning model without actually training it.
Here's the problem the researchers started out with: How does a deep learning model find bugs it's never met before? It'd be like telling someone to identify a unicorn when they'd only ever been exposed to farm cows.
To remedy this data-less training predicament, the researchers cooked up BugLab, wherein two competing deep learning models go head-to-head. One attempts to introduce bugs and hide them in various locations, and the other is tasked with sussing out the little pests its sibling model leaves behind. As a result of the hide-and-seek game, both models get smarter.
If this sounds like a really, really restrained first-act rewrite of Avengers: Age of Ultron, maybe we'll see it turn out to be that way in retrospect. Check back with Windows Central in 2055 when BugLab has determined the ultimate bug is the error in humanity's programming and that the only way to clear the cache on Earth is to wipe us out.
Alternatively, if you're a developer interested in learning about the tech behind Microsoft's bug-hunting operations, check out the official blog post that offers a deep dive into the nitty-gritty of BugLab.
