Future smartphones will be able to understand what you’re taking photos of and recognize faces, says mobile chip maker Qualcomm. Researchers at the company are working to make a powerful new approach to artificial intelligence known as deep learning a standard feature of mobile devices.
Smartphone camera apps often have “scene” modes to get the best shots of landscapes, sports, or sunsets. Qualcomm has created a camera app able to identify different types of scenes on its own, based on their visual characteristics. That could lead to phones that can choose their own settings without having to send or receive data over the Internet.
Charles Bergan, who leads software research at Qualcomm, demonstrated that software in a sponsored talk at MIT Technology Review’s EmTech conference last week in Cambridge, Massachusetts. He said that it should be possible to use the same approach to create software that could decide the best moment to take a photo. “Maybe it will detect that it’s a soccer game and look for that moment when the ball is just lifting off,” he said.
Bergan also demonstrated a facial-recognition app. It recognized his face despite being trained to recognize his features using only a short, shaky, and poorly lit video of his face.
Those demonstrations were based on deep learning, a technique that trains software by processing data through networks of simulated neurons (see “10 Breakthrough Technologies 2013: Deep Learning”). In the case of the scene-classifying app, for example, the simulated neurons were exposed to thousands of photos of different types of scenes.
Bergan said that one reason Qualcomm is working on enabling phones to run deep learning software is that major mobile device manufacturers requested ways to make their devices smarter about images. When exactly the features might make it into phones is unclear.
Qualcomm has previously experimented with chips that are considered “neuromorphic,” because their circuits are arranged in neuron-like arrangements (see “Qualcomm to Build Neuro-Inspired Chips”). However, designs like that are still very much research projects (see “10 Breakthrough Technologies 2014: Neuromorphic Chips”). Bergan says that adding small “accelerators” for deep learning to existing chip designs would be a more practical approach.