Qualcomm has talked about putting “silicon brains” in mobile devices and is now providing tools to train smartphones to recognize people, objects, gestures, and even emotions.
Phones like Samsung’s Galaxy S7 and LG’s G5 that use Qualcomm’s Snapdragon 820 chips will get deep-learning capabilities with the Snapdragon Neural Processing Engine software development kit, announced on Monday.
The SDK will include a run-time that will exploit chip features so smartphones can accomplish deep-learning tasks like tracking objects and recognizing sounds. The kit could also be used in self-driving cars and autonomous drones and robots.
Computers can already recognize people in images, as seen on Facebook. That recognition is thanks to deep-learning systems, which usually require complex neural networks, thousands of servers, and high-performance processors. Qualcomm’s SDK doesn’t need as many resources and is targeted at the tasks and power-consumption levels available to mobile devices.
The SDK is part of Qualcomm’s Zeroth cognitive computing hardware and software platform. In 2013, the company showed a robot with a Zeroth chip and software that was able to make correct navigational decisions on a path after learning and receiving continuous input. Zeroth technology has since been used in mobile devices and is expected to be included in wearable devices, robots, and drones.
For now, the SDK is tuned only for the Snapdragon 820. The chip has digital signal processors that can be programmed to fit deep-learning models. A GPU aids in image processing, and the CPU distributes and schedules tasks.
A similar deep-learning system is already included in Movidius’ thumb PC called Fathom Neural Compute Stick, which uses Google’s TensorFlow machine-learning software to give vision to robots and drones.
Qualcomm’s SDK could expand to industries like security, where surveillance cameras could track and identify a person committing a crime. It could also be used by health care and insurance companies, the company said.
The SDK will support the open-source Caffe deep-learning framework. It will become available in the second half of this year.