However, once the network is trained, the work is much cheaper. Petersen compared his logic gate network to other hyperefficient networks, such as binary neural networks, which use simplified perceptrons that can only process binary values. Logic gate networks perform as well as these other efficient methods in classifying images in the CIFAR-10 data set, which contains low-resolution images in 10 different categories, from “frogs” to “trucks.” I did. It accomplished this in less than one-tenth the logic gates and less than one-thousandth of the time that would otherwise be required. Petersen tested the network using a programmable computer chip called an FPGA. FPGAs can be used to emulate various potential patterns of logic gates. Implementing the network on a non-programmable ASIC chip further reduces costs, as programmable chips require the use of more components to achieve flexibility.
Farinaz Kuchanfar, a professor of electrical and computer engineering at the University of California, San Diego, says he’s not sure logic gate networks can work when faced with more practical problems. “It’s a nice idea, but I don’t know how well it will scale,” she says. She points out that logic gate networks can only be trained approximately through relaxation strategies, and the approximation may fail. That hasn’t caused any problems yet, but Koushanfar says more problems could arise as the network grows.
Nevertheless, Petersen is ambitious. He plans to push the capabilities of logic gate networks even further, eventually hoping to create what he calls a “hardware-based model.” Powerful general-purpose logic gate networks for vision could be mass-produced directly on computer chips, and those chips could be integrated into devices such as personal phones and computers. That could yield huge energy benefits, Petersen said. For example, if these networks can effectively reconstruct photos and videos from low-resolution information, far less data would need to be sent between servers and personal devices.
Petersen acknowledges that logic gate networks will never compete in performance with traditional neural networks, but that’s not his goal. It’s enough to create something that works and is as efficient as possible. “It’s not going to be the best model,” he says. “But it should be the cheapest.”