LOEN, the Lensless Opto-Electronic Neural Network, Cuts the Power Draw of Digit Recognition in Half

Designed to solve the bottleneck of ever-increasing data flows hitting finite computing resources, LOEN uses a mask for convolution.

Gareth Halfacree
2 years ago β€’ Machine Learning & AI

Researchers at Tsinghua University have proposed a solution to the bottleneck issue caused by ever-increasing volumes of data demanding higher and higher power consumption in computer vision systems: a lensless opto-electronic neural network system dubbed LOEN.

"Machine vision faces bottlenecks in computing power consumption and large amounts of data," the researchers explain in the abstract to their paper. "Although opto-electronic hybrid neural networks can provide assistance, they usually have complex structures and are highly dependent on a coherent light source; therefore, they are not suitable for natural lighting environment applications."

LOEN, they claim, is different. The LOEN network is designed to optimize a passive optical mask using a task-oriented neural network, perform optical convolution calculations, and thus reduce both the size of the required device and the computational complexity of the problem β€” with potential for tasks ranging from handwriting recognition to optically-encrypted facial recognition.

It's the mask that makes LOEN notable: designed to replace traditional lenses, the optimized mask modulates the incoming light in order to perform convolution β€” effectively replacing convolutional layers in the neural network running on the host system with a entirely-optical approach, dramatically cutting down on the processing power required.

To prove the concept, the team put together two key experiments. The first used single- and multiple-convolution masks in a hand-written digit recognition network, hitting a classification accuracy of 97.21 percent for the latter while drawing around 50 percent less energy than a traditional implementation.

The second experiment took the concept a stage further, using the masks to perform convolution on images of individuals such that they are entirely unrecognizable to a human observer but can still be fed through a facial recognition system β€” optical encryption, in effect. In this experiment, the researchers report effectively the same recognition accuracy as facial recognition networks running on unencrypted imagery.

The team's work has been published under open-access terms in the journal Light Science & Applications.

Gareth Halfacree
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles