As a result, thermodynamic computing has recently become a novel and promising approach to lowering energy costs in AI image generation. Researchers, led by Stephen Whitelam at the Lawrence Berkeley National Laboratory, have developed a method that may allow thermodynamic computers to generate images using just one ten billionth the energy of current digital methods. This finding represents a significant breakthrough in the field of machine learning and image analysis.
In a study published on January 10 in Nature Communications, Whitelam and a colleague presented their findings on creating a thermodynamic version of neural networks. This cutting-edge approach includes teaching the machine with a carefully curated collection of images. As a result, the computer can then output completely new images—like these handwritten digits.
Whitelam emphasizes the potential of thermodynamic computing, stating that it could produce images “with a much lower energy cost than current digital hardware can.” This is especially important in light of the increased scrutiny on the environmental effects of energy-intensive computing workflows, like artificial intelligence.
Thermodynamic computers are even earlier in their training. Compare that to conventional, digital neural networks, which companies today are already implementing on massive scale. To their delight, Whitelam’s simulations—published January 20, 2022 in Physical Review Letters—produced groundbreaking results for each simulation. Well, these folks showed that these computers can create realistic photos of handwritten feet! This result shows that for certain applications, thermodynamic computing could one day outperform today’s technologies.
One notable development in this field is the prototype chip created by Normal Computing, a startup based in New York City. The monolithic chip consists of eight different resonators connected by their own custom couplers. This design marks the technology’s nascent stage as a thermodynamic computer vs being a digital computer. Whitelam points out that unlike conventional vowel recognition systems, these devices do not rely on energy-intensive digital neural networks. They avoid the use of noise-generating pseudorandom number generators, further upping their energy efficiency.
Whitelam readily admits that there’s tough work yet to come. “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E,” he stated. He added, “It will still be necessary to work out how to build the hardware to do this.”
Whitelam and his team are using this initial work to further investigate how thermodynamic computing could be used to generate images. Their findings point to a future where we can realize the full potential of machine learning at much lower energy expenditures than today. As research continues, this technology could truly have game-changing implications. It will be a paradigm shift across industries that need advanced computing, well beyond its impact on image generation.