Scientists have developed a novel "thermodynamic computer" that generates images by harnessing random energy fluctuations, known as thermal noise. This new system mimics the capabilities of artificial intelligence (AI) neural networks but achieves this with orders of magnitude less energy.
Unlike conventional AI, which generates images using chips where energy expenditure to flip bits is significant, this "generative thermodynamic computer" leverages system noise. Researchers outlined their findings in the journal Physical Review Letters. The concept draws an analogy to surfing, where wave power is harnessed, contrasting with an ocean liner plowing through waves expensively.
This approach is a form of probabilistic computing, utilizing thermal noise fluctuations to power computations. This efficiency is particularly beneficial for optimization problems. Researchers at Normal Computing Corporation have built circuits operating at low energies comparable to thermal noise. These circuits can be programmed to pose questions that the resulting equilibrium fluctuations can answer, such as solving linear algebra problems.
The thermodynamic computer can generate images by reversing the process of an image being obscured by noise. By manipulating equations like the Langevin equation, it calculates probabilities to reconstruct images from random noise, potentially creating novel images not present in the training data. Experts describe the findings as important, highlighting the growing interest in probabilistic computing and its potential to offer fundamental interpretations in fields often dominated by "black-box" AI models.