Stephen Whitelam is a staff scientist at Lawrence Berkeley National Laboratory in California. He’s clearly done something amazing in AI by riding the new, generative wave of computing. Along with a colleague, Whitelam published a groundbreaking article in Nature Communications on January 10, which reveals the potential of creating a thermodynamic version of a neural network.
Limitation though they are, compared to their digital counterparts, thermodynamic computers hold thrilling potential. Whitelam’s research found that, with the right training process, a thermodynamic computer would be able to create pictures of handwritten digits. You trained them entirely through simulations on standard, off-the-shelf computers. It was a followup publication in Physical Review Letters on January 20 that filled in the details.
Perhaps the most astonishing result from Whitelam’s study is the energy efficiency of thermodynamic computing. Its environmental impact is staggering. It can produce pictures with just one ten billionth of the power that our current digital neural networks require. Such unbelievable efficiency must breed innovation, right? This radical decrease in energy consumption is not based on the energy-intensive processes that we have come to associate with digital computing. Thermodynamic computing removes the requirement for noise-creating pseudorandom number generators, increasing its efficiency.
Normal Computing, a New York City-based startup, created a prototype chip. This ground-breaking chip consists of eight resonators linked by customized couplers. Whitelam has suggested a strategy that might help this thermodynamic computer create an impressive sequence of images.
Whitelam emphasizes the significance of this research, 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.” He agrees that the idea is promising, but admits there’s still a lot to be learned about the design and potential of thermodynamic computers. He cautions, “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E.”
This new research has the potential to shake up the field of AI-image generators. Which makes it a very attractive solution for developers and researchers. According to Whitelam, thermodynamic computing might indeed generate images “with a much lower energy cost than current digital hardware can.”