Chip Makers Adding Brains And Eyes To AI
Chip Makers Adding Brains And Eyes To AI.
Chip companies are adding greater smarts to cameras, spurring a new generation of equipment that not only captures imagery but interprets and acts on what it sees. Such advances in computer vision—the ability to extract information from images—can enable, say, a network of security cameras to track a package’s movement. Or, in the case of Apple Inc.’s newly unveiled iPhone X, unlock a smartphone by recognising a person’s face. Alphabet Inc.’s Nest Labs in September announced a doorbell equipped with a Qualcomm Inc. chip, a video camera and facial-recognition software that can send an alert to a Nest mobile app if it sees a familiar face. The market for computer-vision systems is nascent, poised to expand from roughly $1 billion last year to $2.6 billion in 2021, according to International Data Corp. Emerging products such as autonomous vehicles and personal robots portend continuing growth, and Intel Corp., Qualcomm and other chip makers are jockeying to supply the brains to new machines. “These [applications] are edging into viability,” said IDC analyst Michael Palma. “Maybe not mass viability, but very, very close.”
Blue River Technology, a Silicon Valley startup acquired for $305 million last month by Deere & Co., is using computer vision powered by Nvidia Corp. to help lettuce farmers boost productivity and reduce or reallocate labor costs. Farmers tend to plant lettuce seeds densely and then thin the overcrowded sprouts using hoes, a time-consuming operation. Blue River’s See & Spray, a rig that hitches to the back of a tractor, uses up to two dozen cameras, each equipped with an Nvidia computer called Jetson, to identify individual sprouts and evaluate their distance from neighbours with quarter-inch accuracy. Those too close together get doused automatically with a precisely aimed shot of fertiliser, enough to kill an individual plant even as it nourishes the field—no manual labor required. See & Spray can typically thin an acre of lettuce in 12 minutes, work that would take a person eight hours, according to Richard Smith, a specialist in vegetable crop production from the University of California, Davis. Blue River claims the machine can increase yields by 10%. Deere plans to extend the technology to other crops as part of its effort to shift agriculture from tending fields to nurturing individual plants. Willy Pell, who oversees new technology at Blue River, believes machines outfitted to perceive the world and act on what they sense without human intervention will drive the next wave of Silicon Valley investment. “There’s a lot of humanity that’s simply using eyes and hands to do things,” he said. Machines outfitted with camera eyes and silicon brains soon will be able to take over “all kinds of repetitive tasks.”
The same technology is bringing new capabilities to consumer products as well. Qualcomm’s next-generation Snapdragon smartphone chips will transform camera output into detailed 3-D maps, a boon for superimposing computer-generated imagery over real-world scenes in augmented-reality apps. Myriad, a line of chips from Intel’s Movidius division that performs artificial-intelligence computations using very little electrical power, has found a niche in security cameras and drones, and is branching into medicine. Doctor Hazel, a startup, created an AI tool using a Myriad chip that works with a medical camera to detect skin cancers on the spot. It diagnoses cancers with up to 85% accuracy, and that rate should improve as the system is further trained with images of known benign and malignant moles, according to Doctor Hazel co-founder Mike Borozdin. One advantage to these new computer-vision systems: They pack enough computing horsepower to apply AI to images locally, rather than needing to interact with remote servers. That speeds up processing, enabling devices to work without a reliable network connection—for, say, a drone inspecting turbines on a wind farm—while avoiding the risk of exposing information that may be private or proprietary.
Deepu Talla, Nvidia’s vice president in charge of AI for applications such as robotics and drones, believes both local and remote processing will be necessary. Cameras mounted on traffic lights in an urban area can, for instance, count passing vehicles and forward their tallies to a cloud data center that analyses the output and controls the lights to keep traffic flowing smoothly. Nvidia is working on similar systems in Hangzhou, China, with Alibaba Group Holding Ltd. and in Shenzhen, China, with Huawei Technologies Co. “There’s not enough human eyeballs available” to make sense of all the imagery cameras will capture, said Mr. Talla. “You need computer vision.”
Credit: Ted Greenwald for The Wall Street Journal 4 October 2017.
