Stephen Whitelam is a staff scientist at Lawrence Berkeley National Laboratory in California. In this interview, he provides a look at the impact he’s had with his innovative work in thermodynamic computing. His latest research demonstrates that this novel technology is able to synthesize the same images as current digital hardware using one ten billionth of the energy. This innovative advancement could truly transform the industry. This innovative approach could significantly change the way AI generates images. It would make a big dent in the energy footprint associated with the digital neural networks.
On January 10, Whitelam and a colleague published that original article in Nature Communications on this fascinating, illuminating story. They explained their novel proposal of developing a thermodynamic implementation of a neural network. The researchers identified a novel training procedure that trains a thermodynamic laptop to produce photos of handwritten digits. It’s a creative model that actually works super well. We begin training the machine by providing it with a large dataset of images. This necessary step gives the magic cat all the information he needs to generate original graphics.
Whitelam explained, “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.” This announcement highlights the potential of thermodynamic computing to spur innovations in AI image generation while preserving our planet’s precious energy resources.
All digital neural networks in operation today are energy-hungry and shoot the bulk of their neurons to noise through pseudorandom number generators. Whitelam’s approach would make this process a lot easier. He stressed that thermodynamic computers are still rather basic compared to their digital cousins, but they all the same mark an important breakthrough.
Whitelam and his group at Normal Computing have created a groundbreaking prototype chip. The other side of the chip shows eight resonators interconnected through specialized couplers. Though still an early design that is just being tested, researchers concede there are challenges to overcome. Whitelam remarked, “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E,” highlighting the gap between current capabilities and potential future applications.
The research describing these advances was published Jan. 20 in Physical Review Letters. However, to elucidate these ideas more thoroughly, the researchers simulated a thermodynamic computer on standard machines to further explore its potential capabilities. Whitelam underscored that though there is great promise at a theoretical level, practical development of the hardware is key to moving the idea into real-world uses. He stated, “It will still be necessary to work out how to build the hardware to do this.”