Godfather of AI Geoffrey Hinton and John J Hopfield Win Nobel Prize in Physics

Geoffrey E Hinton and John J Hopfield were awarded the Nobel Physics Prize 2024 in Physics by The Royal Swedish Academy of Sciences for pioneering advancements that form the basis of today’s ML world. Their work used principles of physics to develop neural networks.

Hinton, often referred to as the ‘Godfather of AI,’ achieved the award on the basis of developing the ‘Boltzmann machine’, a neural network model inspired by statistical physics. This machine allows neural networks to self-learn patterns from data by modelling systems with interacting nodes, mimicking how the brain processes and categorises information. This technique is foundational for technologies like image and speech recognition.

Hopfield grabbed the Nobel by approaching associative memory and pattern recognition, which involved creating the ‘Hopfield network,’ that applies physics concepts to create a system where the energy configuration is optimised to recreate stored images or data. This has influenced numerous applications in AI.

The network is described similarly to the energy of a spin system in physics, and it is trained by adjusting the connections between nodes to minimise the energy of stored images. When given a distorted or incomplete image, the Hopfield network systematically updates the node values to reduce the network’s overall energy. Through this step-by-step process, it retrieves the stored image that most closely resembles the imperfect input.

“The laureates’ work has already been of the greatest benefit. In physics we use artificial neural networks in a vast range of areas, such as developing new materials with specific properties,” said Ellen Moons, chair of the Nobel Committee for Physics.

Hinton’s most cited paper, ‘Image Net Classification with Deep Convolutional Neural Networks’ discusses the use of Convolutional Neural Networks (CNNs) for image classification, specifically on the ImageNet dataset which contains over 15 million labelled high-resolution images. The authors trained one of the largest CNNs to date on subsets of the ImageNet dataset used in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC).

Hinton has previously introduced ‘Forward Forward’ (FF) algorithm as an alternative to traditional backpropagation for training neural networks. Unlike backpropagation, which adjusts weights in a backward pass, the FF algorithm uses two forward passes to manage weights, one increasing goodness for correct data and one reducing it for incorrect data. 

This mimics the human brain’s processing and allows for lighter, faster training with less computational power. Hinton believes this could lead to more efficient, brain-like AI systems.