Google Street View is brilliant. It finds us when we’re lost, it shows us where we are, it reveals places we’ll never get to visit, and so on and so forth. But you know what’s even more amazing? The crazy neural network that Street View is built on.
Take, for example, the conundrum of house numbers in panoramic images, like the ones Street View shows you. There hundreds of millions of street numbers in a single country alone — so finding and identifying each number in Street View would be a sisyphean task. Instead, Google engineers built a “deep convolutional neural network” that operates on the pixels of the images themselves, combing them for number imagery and logging what it finds.
This neural network — which you can read about here, basically it’s a computing network modelled on animal nervous systems — has eleven layers of neurons, which makes it possible to ID millions of house numbers a day from the Street View raw image data. “We can, for example, transcribe all the views we have of street numbers in France in less than an hour using our Google infrastructure,” write the engineers in a new Arxiv paper about the project.
What about the numbers that are too blurry for this giant brain to make sense of? No prob — those are identified by humans as part of a second generation CAPTCHA program. So you may have already contributed to the cause, without even realising it. [Arxiv; MIT Technology Review]