The next generation of neural networks may reside in hardware

Once the network is trained, however, things get cheaper. Peterson compared his logic-gate network to a set of other high-performance networks, such as binary neural networks, which use simple perceptrons that can only process binary values. Logic-gate networks did this similarly to other efficient methods for classifying images in the CIFAR-10 data set, which includes 10 different categories of low-resolution pictures ranging from “frog” to “truck.” He achieved this in less than one-tenth and less than one-thousandth the number of logic gates required by those other methods. Peterson tested his network using programmable computer chips called FPGAs, which could be used to simulate various possible patterns of logic gates; Implementing the network in non-programmable ASIC chips will further reduce costs, as programmable chips need to use more components to achieve their flexibility.

Farinaz Kaushanfar, a professor of electrical and computer engineering at the University of California, San Diego, says he’s not sure logic-gate networks will perform when faced with more realistic problems. “It’s a beautiful idea, but I’m not sure how well it scales,” she says. She notes that logic-gate networks can only be trained approximately by relaxation strategies, and approximations can fail. That hasn’t caused problems yet, but Kaushanfar says it could prove more problematic as the network grows.

Still, Peterson is ambitious. He plans to continue to push the capabilities of his logic-gate networks, and he hopes to eventually create what he calls a “hardware foundation model.” A powerful, general-purpose logic-gate network for vision can be mass-produced directly on computer chips, and those chips can be integrated into devices such as personal phones and computers. That can have enormous energy benefits, says Peterson. If those networks could efficiently reconstruct photos and videos from low-resolution information, for example, much less data would need to be sent between servers and personal devices.

Peterson admits that logic-gate networks will never compete with traditional neural networks on performance, but that is not his goal. Building something that works, and is as efficient as possible, should be enough. “It won’t be the best model,” he says. “But it has to be the cheapest.”