Stephen Whitelam, a staff scientist at Lawrence Berkeley National Laboratory in California, has introduced an innovative approach to image generation using thermodynamic computing. This new approach provides a more energy-efficient and digital alternative to conventional ANN. It would lower the energy expended for image production to only one ten billionth of what today’s techniques use.
Whitelam’s approach involves a thermodynamic computer that goes through an initial set of images. This technology is still nascent compared to more established digital processes, its potential is remarkable. Previously, research has demonstrated that thermodynamic computers can perform complex tasks such as image generation. They accomplish all this while using about three-quarters less energy. Whitelam emphasizes the potential of this technology by stating, “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.”
The conceptual underpinnings of Whitelam’s study were published in the journal Nature Communications on Jan. 10. He and his coauthor created a thermodynamic-inspired neural network. This state-of-the-art generative model is capable of creating extraordinarily realistic images, such as handwritten digits, as it learns. New York City-based startup Normal Computing has created a groundbreaking prototype chip that uses a completely different approach. It features eight resonators, which are linked together through custom couplers.
Despite all the progress, Whitelam admits there’s a long road ahead. He notes, “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E.” That’s why he says we need to find a better way to manufacture that hardware. This second step is very important towards producing the intended results in text to image generation.
Unlike their digital counterparts, thermodynamic computers do not need energy-hogging neural networks or noise-producing pseudorandom number generators. This key difference might unlock a new generation of energy-efficient computing solutions.