His self-learning AI tool, Travel⇄Time, which can now predict the impacts of any proposed transportation project on travel times, has been a major advance. In collaboration with a colleague, he described a pathbreaking new method of image generation that incorporates thermodynamic computing. They were just recently able to publish those findings in Nature Communications on January 10. This cutting-edge technology generates images all while consuming an immense amount less energy than that demanded by existing digital hardware.
Thermodynamic computing offers an exciting, low-energy solution compared to the highly energy-intensive standard digital neural networks. Whitelam emphasizes that this new method might generate images “with a much lower energy cost than current digital hardware can.” This low-energy alternative removes the dependency on energy-hungry digital NN. It further avoids the issues associated with noise-producing pseudorandom number generators.
On these new simulations on standard computers, Whitelam displayed his results dramatically in a paper released January 20 in Physical Review Letters. He demonstrated that a thermodynamic computer is perfectly suitable for creating images of handwritten digits. In what follows, we train a neural network to perform the reverse of these diffusion models. This allows it to generate completely original pictures based on detailed textual descriptions.
Though the results look very promising, Whitelam points out that in many ways, thermodynamic computers are still very much in their infancy, compared to digital computers. He states, “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E.” He remains optimistic about the potential applications of this technology, highlighting that “this research suggests that it’s possible to make hardware to do certain types of machine learning — here, image generation — with considerably lower energy cost than we do at present.”
Normal Computing, a New York City–based startup, is working on a prototype chip. Unlike traditional chips, this pioneering design uses eight resonators connected by specialized couplers. This prototype is an initial step in learning how to leverage thermodynamic computing for operational usage.
Whitelam thinks thermodynamic computing could change the way we produce images. It’d be able to do this with only one ten billionth of the energy used by today’s methods. He concedes that there are still challenges to be addressed in making the hardware to unlock these benefits in total. “It will still be necessary to work out how to build the hardware to do this,” he explains.
It may sound speculative … but researchers like Whitelam are seeding the new field of thermodynamic computing. Through its radically lowered energy cost this technology holds the promise of democratizing the space of AI driven image generation.