Startup aims to improve efficiency of deep neural networks for image analysis and facial recognition

Purdue University-affiliated startup FWDNXT has developed a low-power mobile coprocessor for accelerating image recognition and classification by deep neural networks, Purdue News reports.

The accelerator is called Snowflake, and running on a Field Programmable Gate Array module it can achieve 91 percent computational efficiency on entire convolutional neural networks, with 99 percent efficiency on some layers. FWDNXT says its expertise in scene analysis and scene parsing could make its hardware and software useful to autonomous vehicles, facial recognition, and day-to-day uses like identifying items from a shopping list.

The company, which is based in Purdue Research Park, originally developed its technology with grants from Purdue and the Navy, and has now obtained multimillion dollar funding through a strategic partnership. It is now planning to add to its team, and pursue Series A funding to fuel its ambition to make microchips used in virtually all smart devices.

“Everybody was looking for a solution like this,” said Eugenio Culurciello, an associate professor at Purdue’s Weldon School of Biomedical Engineering. “We have a special computer that can operate on large data very fast with low power consumption. Our mission is to propel machine intelligence to the next level.”

FWDNXT has filed patent applications through the Purdue Office of Technology Commercialization. It plans to complete the development of a prototype microchip in the first half of 2018, and hopes to be able to sell programable logic prototypes soon.

As previously reported, a Purdue University professor recently developed a technology for 3D image transmission with possible applications for facial recognition.

Article Topics

image recognition  |  neural networks  |  Purdue  |  research and development