Convolutional neural network on-a-chip promises always-on face recognition

June 15, 2017 // By Julien Happich
Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a Convolutional Neural Network Processor on silicon that they combined with a custom made image sensor to perform face recognition with a 97% accuracy while only drawing 0.62mW.

Using such embedded artificial intelligence, the researchers claim their solution consumed only 1/5000 the power that would be required by a GPU performing the same tasks.

This unique processor owes its incredibly low power consumption thanks to its CNN whose circuitry, architecture, and algorithms have all gone through optimization steps. On-chip memory has been integrated in the CNNP so it could be read in a vertical direction as well as in a horizontal direction, reports KAIST. In the CNNP, 1024 multipliers and accumulators operate in parallel and the chip is capable of directly transferring the temporal results to any of those without access to external memory or to an on-chip communication network. The chip which was presented at the International Solid-State Circuit Conference (ISSCC) held in San Francisco last February performs convolution calculations with a two-dimensional filter in the CNN algorithm, approximated into two sequential calculations of one-dimensional filters to achieve higher speeds and lower power consumption.
On the basis of this chip developed by Kyeongryeol Bong, a Ph. D. student under Professor Hoi-Jun Yoo of the Department of Electrical Engineering, and in collaboration with Korean start-up UX Factory Co., the researchers developed a wearable face recognition system they hope to bring to market by the end of the year.