Yesterday, we received some exciting news: the Nobel Prize in Physics was awarded to John Hopfield and Geoffrey Hinton, two pioneers of artificial intelligence, “for foundational discoveries and inventions that enable machine learning with artificial neural networks”.
Beyond the usual speculation accompanying the Nobel Prize predictions, several comments have surfaced questioning the prize’s decision, suggesting that AI is being awarded “because it’s trendy” rather than for its relevance to physics. On the contrary, we believe the prize is more than justified. At a fundamental level, much of existing AI is inspired by physical phenomena, and it is precisely this inspiration that makes these methods so powerful.
Specifically, John Hopfield, inspired by spin systems, created a network capable of storing and regenerating patterns, employing the physical dynamics of spin systems and attempting to configure their interconnections in such a way that the images stored in the network reach the lowest possible energy, in the physical sense of the term. In this way, if the network receives a distorted or incomplete image, such as a damaged or noisy one, it can modify it according to the pictures it has previously seen and return the most complete version that matches the input. As incredible as it seems, this entire process is achieved solely using the laws of spin physics, treating each spin like a small magnet and using their magnetic interactions for something as seemingly unrelated as pattern recognition.
Building on this work, Geoffrey Hinton developed a type of spin network called a Boltzmann machine, which uses a physical approach different from Hopfield’s network. In this case, the network is designed to identify patterns after training with expected events (just like a traditional neural network). This mechanism made it an optimal choice for classification and synthetic data generation, and it remained state-of-the-art until the rise of other architectures, such as GANs.
At Quside, on a fundamental level, we work with models very similar to those used by Hopfield and Hinton. For a long time, we have been working with spin systems, exploring how the simulation of these systems can be affected by the quality of the random numbers used to simulate them and developing new methods to find those minimum energy states more efficiently by leveraging the existence of a high-speed entropy source. Moreover, together with our collaborators, we explore how Boltzmann networks can be used to simulate random distributions with a degree of complexity too high to be modeled by other, less versatile methods.
Awards like this, beyond recognizing how physics has become a multidisciplinary science that permeates and penetrates all fields, help us advance basic research, both in its own right and through its applications at the scientific, economic, and social levels.
Our sincerest congratulations to the laureates and all those who believe that physics and AI have an incredible journey ahead, hand in hand.
