Once the network has been trained, though, things get way, way cheaper. Petersen compared his logic-gate networks with a cohort of other ultra-efficient networks, akin to binary neural networks, which use simplified perceptrons that may process only binary values. The logic-gate networks did just in addition to these other efficient methods at classifying images within the CIFAR-10 data set, which incorporates 10 different categories of low-resolution pictures, from “frog” to “truck.” It achieved this with fewer than a tenth of the logic gates required by those other methods, and in lower than a thousandth of the time. Petersen tested his networks using programmable computer chips called FPGAs, which will be used to emulate many alternative potential patterns of logic gates; implementing the networks in non-programmable ASIC chips would cut back costs even further, because programmable chips need to make use of more components with a view to achieve their flexibility.
Farinaz Koushanfar, a professor of electrical and computer engineering on the University of California, San Diego, says she isn’t convinced that logic-gate networks will find a way to perform when faced with more realistic problems. “It’s a cute idea, but I’m undecided how well it scales,” she says. She notes that the logic-gate networks can only be trained roughly, via the comfort strategy, and approximations can fail. That hasn’t caused issues yet, but Koushanfar says that it could prove more problematic because the networks grow.
Nevertheless, Petersen is ambitious. He plans to proceed pushing the skills of his logic-gate networks, and he hopes, eventually, to create what he calls a “hardware foundation model.” A strong, general-purpose logic-gate network for vision might be mass-produced directly on computer chips, and people chips might be integrated into devices like personal phones and computers. That would reap enormous energy advantages, Petersen says. If those networks could effectively reconstruct photos and videos from low-resolution information, for instance, then far less data would have to be sent between servers and private devices.
Petersen acknowledges that logic-gate networks won’t ever compete with traditional neural networks on performance, but that isn’t his goal. Making something that works, and that’s as efficient as possible, must be enough. “It won’t be the very best model,” he says. “Nevertheless it must be the most affordable.”
