SAN JOSE, Calif. — Lab work is extending machine learning to serve new applications and define new hardware architectures, said a Qualcomm researcher. He spoke on the occasion of the company acquiring Scyfer B.V., a small AI research team affiliated with University of Amsterdam that it had been working with previously.
Scyfer acted as a consulting firm, applying machine learning to industrial, IoT, banking, and mobile sectors. The group is now part of Qualcomm Research, seeking to expand machine learning in areas such as computer vision and natural language processing and exploring how emerging algorithms will impact the design of hardware accelerators.
“As the algorithms change, we think there is a space here for co-designing the neural networks and the hardware,” said Jeff Gehlhaar, a vice president of technology for corporate R&D who is responsible for AI at Qualcomm.
“As these networks evolve, we are starting to see patterns in execution profiles” that can impact caching and bit precision in hardware, he said. “We’re looking at the systems level to see how to make elements work together without compromising accuracy.”
In a press statement, Qualcomm said that it foresees emerging uses of neural nets in areas including wireless connectivity, power management, and photography. Several companies are already applying AI to security for jobs such as detecting malware; others are using it to break through bottlenecks in semiconductor design.
Although neural nets have been a research focus for decades, 2013 was a watershed year, with AlexNet results giving birth to the deep-learning field. Click to enlarge. (Image: Qualcomm)
Qualcomm’s challenge is in trying to move more AI workloads out of server farms in the cloud and onto its chips in smartphones. To that end, it released last year a neural network developer’s kit for its Snapdragon SoC. It partnered with Google and Facebook to optimize it for their TensorFlow and Caffe 2 frameworks.
The Amsterdam team will also boost Qualcomm’s work on generative adversarial networks, an emerging algorithm that the company thinks could drive more training jobs to handsets. Machine learning in the cloud “will continue to be dominant, [but] we see a great opportunity and customer requests to do versions of it on devices” for jobs such as face recognition, said Gehlhaar.
“The trend to deploy AI on edge devices like drones, cars, and smartphones is relatively new, but our customers are already doing this — it’s not lab experiments; people are solving real-world problems.”
Machine learning “is a rapidly evolving field,” said Gehlhaar, noting that advances in image recognition in 2013 with AlexNet spawned the deep-learning field that is now his group’s main focus. A recent panel of experts underscored the rapidly rising importance of the field as well as its limits.
— Rick Merritt, Silicon Valley Bureau Chief, EE Times