Stephen Whitelam, a staff scientist at Lawrence Berkeley National Laboratory in California, has done extraordinary work to thermodynamic computing. He has shown us that we can do a thermodynamic implementation of a version of a neural network. This innovation could potentially transform image generation processes, allowing for the creation of images using a fraction of the energy consumed by current digital hardware.
In a recent publication in Nature Communications on January 10, Whitelam and a colleague demonstrated that the training process can lead to a thermodynamic computer capable of generating images of handwritten digits. This novel approach presents a lighter, more efficient, and energy-saving solution compared to conventional digital neural networks. Unlike those overly energy-intensive models, it doesn’t use pseudorandom number generators that spit out noise.
Whitelam’s work in thermodynamic computing has the potential to change the landscape of image generation. of the energy currently consumed by digital systems, potentially lowering consumption to one-tenth billionth. He emphasizes the potential of this technology, 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.”
Though these are still encouraging signs, Whitelam is quick to note that thermodynamic computers remain elementary beside their digital siblings. 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.” Considerable progress will be required to produce hardware that can get anywhere near the performance of well-tuned digital systems.
To showcase this notion, Whitelam’s latest approach consists of training a thermodynamic pc utilizing a library of images. Normal Computing has created their own prototype chip, which includes eight resonators. These resonators couple via custom couplers, showing the real-world implementation of this cutting-edge technology.
Whitelam’s findings were shared in simulations published in Physical Review Letters on January 20, further solidifying the groundwork for future research in this area. He emphasizes that while the potential energy savings are substantial, “It will still be necessary to work out how to build the hardware to do this.”
