As reported by MIT Technology Review: Smartphone camera apps could soon do more than just capture images. Software integrated with a new line of smartphone chips will be capable of recognizing, say, a cat or city skyline in a photo, and tagging pictures of your friends with their names.
The chip maker Qualcomm announced last week that it will bundle the software with its next major chip for mobile devices. The technology, which Qualcomm calls Zeroth, could make sophisticated machine learning more common on mobile devices. As well as processing images, the Zeroth software is designed to allow phones to recognize speech or other sounds, and to learn to spot patterns of activity from a device’s sensors.
The technology uses an approach to machine learning known as deep learning that has led to recent advances in speech and object recognition, as well as software able to play Atari games with superhuman skill (see “Google’s Intelligence Designer”). Deep learning software is loosely modeled on some features of brains. It can be trained to recognize certain objects in images by processing many example photos through a network of artificial “neurons” arranged into hierarchical layers.
Tim Leland, a vice president of product management with Qualcomm, says the company plans to work with partners to build apps to make use of the new capabilities. He wouldn’t say exactly what those apps might do, but Qualcomm’s demonstrations of the technology have so far focused mostly on enhancing the features of camera apps. Last week at the Mobile World Congress event in Barcelona, where Zeroth was announced, Qualcomm showed it powering a camera app that could recognize faces it had seen before, and detect different types of photo scenes.
The Zeroth software is being developed to launch with Qualcomm’s Snapdragon 820 processor, which will enter production later this year. The chip and the Zeroth software are also aimed at manufacturers of drones and robots.
Normally an app has to send data out over the Internet to a powerful server in order to perform such tricks (see “10 Breakthrough Technologies 2013: Deep Learning”). By doing such computation on the phone itself, the software should be better at interpreting data from location and motion sensors on a device, Leland says. He predicts that one of the first applications of Zeroth will be extending the battery life of devices by tracking the way a person uses a phone and learning when it could power down to save energy without affecting the user experience.
Qualcomm is also experimenting with chips that have physical networks of “neurons” made from silicon that communicate with spiking electrical signals (see “Qualcomm to Build Neuro-Inspired Chips”). That might give devices more powerful learning abilities, but it would mean they’d work much differently than today’s devices do.