The next generation of neural networks could live in hardware

However, once the network is trained, things become cheaper. Peterson compared his network of logic gates to a set of other ultra-efficient networks, such as binary neural networks, which use simplified perceptrons that can only process binary values. Gatenet performed as well as other effective methods when classifying images in the CIFAR-10 dataset, which includes 10 different categories of low-resolution images ranging from “frog” to “truck.” It does this in less than one-tenth the logic gates required by other methods and in less than one-thousandth the time. Petersen tested his network using a programmable computer chip called an FPGA, which can be used to simulate many different potential modes of logic gates; implementing the network in a non-programmable ASIC chip would further reduce costs, since programmable chips require Its flexibility is achieved by using more components.

Farinaz Koushanfar, a professor of electrical and computer engineering at the University of California, San Diego, said she doesn’t believe networks of logic gates can work when faced with more real-world problems. “It’s a lovely idea, but I’m not sure how well it scales,” she said. She pointed out that logic gate networks can only be trained approximately through relaxation strategies, and the approximation may fail. This hasn’t caused problems yet, but Koshanfar said it could cause more problems as the network grows.

Still, Peterson is ambitious. He plans to continue to improve the capabilities of the logic gate network and hopes to eventually create what he calls a “hardware base model.” Powerful general-purpose logic gate networks for vision can be mass-produced directly on computer chips, and these chips can be integrated into devices such as personal phones and computers. Peterson said this could lead to huge energy benefits. For example, if these networks could efficiently reconstruct photos and videos from low-resolution information, much less data would need to be sent between servers and personal devices.

Peterson acknowledges that logic gates will never compete with traditional neural networks in terms of performance, but that’s not his goal. Making something that works and is as efficient as possible is enough. “It’s not going to be the best model,” he said. “But it should be the cheapest.”