Nobel Prize in Physics 2024

John Hopfield and Geoffrey Hinton pioneered computational methods that enabled the development of neural networks.

The backpropagation of error (backprop) algorithm is frequently used to train deep neural networks in machine learning, but it has not been viewed as being implemented by the brain. In this Perspective, however, Lillicrap and colleagues argue that the key principles underlying backprop may indeed have a role in brain function.

Dmitry Krotov discusses recent theoretical advances in Hopfield networks and their broader impact in the context of energy-based neural architectures.

Although the initial inspiration of neural networks came from biology, insights from physics have helped neural networks to become usable. New connections between physics and machine learning produce powerful computational methods.

Present day quantum technologies enable computations with tens and soon hundreds of qubits. A major outstanding challenge is to measure and benchmark the complete quantum state, a task that grows exponentially with the system size. Generative models based on restricted Boltzmann machines and recurrent neural networks can be employed to solve this quantum tomography problem in a scalable manner.

Neuromorphic computing might be the answer to AI’s hardware problem.

A type of stochastic neural network called a restricted Boltzmann machine has been widely used in artificial intelligence applications for decades. They are now finding new life in the simulation of complex wavefunctions in quantum many-body physics.

There is still a wide variety of challenges that restrict the rapid growth of neuromorphic algorithmic and application development. Addressing these challenges is essential for the research community to be able to effectively use neuromorphic computers in the future.

Powered by Nature Navigator

Brains and neuromorphic systems learn with local learning rules in online-continual learning scenarios. Designing neural networks that learn effectively under these conditions is challenging. The authors introduce a neural network that implements an effective, principled approach to local, online-continual learning on associative memory tasks.

Learning-to-learn refers to progressive speedup in solving a series of problems with shared structure. This study shows that it emerges in recurrent neural networks from the reuse and refinement of a neural state subspace underlying schema formation.

Memristors are passive electrical components that can act like simple memories. Here, the authors use an array of hafnium oxide memristors to create a type of artificial neural network, known as a Hopfield network, that is capable of retrieving data from partial information

This paper introduces ‘prospective configuration’, a new principle for learning in neural networks, which differs from backpropagation and is more efficient in learning and more consistent with data on neural activity and behavior.

The authors derive a neural network theory of systems consolidation to assess why some memories consolidate more than others. They propose that brains regulate consolidation to optimize generalization, so only predictable memory components consolidate.

The authors present DPAD, a deep learning method, for dynamical neural–behavioral modeling. It dissociates behaviorally relevant neural dynamics, better predicts neural–behavioral data and reveals insight into where their nonlinearities can be isolated.

The authors identify reusable ‘dynamical motifs’ in artificial neural networks. These motifs enable flexible recombination of previously learned capabilities, promoting modular, compositional computation and rapid transfer learning. This discovery sheds light on the fundamental building blocks of intelligent behavior.

This Review provides an overview of memory devices and the key computational primitives for in-memory computing, and examines the possibilities of applying this computing approach to a wide range of applications.

Phase-change memtransistive synapses enable the implementation of biomimetic neural algorithms to perform tasks such as sequential learning and stochastic Hopfield computing networks.

Stochastic orbital dynamics of individually coupled Co atoms on black phosphorus enables the realization of a Boltzmann machine capable of self-adaption.

This study shows that by enhancing internal complexity of neurons in a Hodgkin–Huxley network, similar performance to larger, simpler networks can be achieved, suggesting an alternative path for powerful AI systems by focusing on neuron complexity.

Memristors are devices that possess materials-level complex dynamics that can be used for computing, such that each memristor can functionally replace elaborate digital circuits. This Review surveys novel material properties that enable complex dynamics and new computing architectures that offer dramatically greater computing efficiency than conventional computers.

Memristors hold promise for massively-parallel computing at low power. Aguirre et al. provide a comprehensive protocol of the materials and methods for designing memristive artificial neural networks with the detailed working principles of each building block and the tools for performance evaluation.

As physicists are increasingly reliant on artificial intelligence (AI) methods in their research, we ponder the role of human beings in future scientific discoveries. Will we be guides to AI, or be guided by it?

Minimizing the energy of the Ising model is a prototypical combinatorial optimization problem, ubiquitous in our increasingly automated world. This Review surveys Ising machines — special-purpose hardware solvers for this problem — and examines the various operating principles and compares their performance.

Variational solutions to correlated open quantum systems rely on an efficient parametrization of the density matrix. In this work the authors show how deep neural networks can be used to obtain good approximations of steady-state density matrices without the restriction of being positive.

Neural network quantum states (NQS) are a promising method to simulate large fermionic systems. This work reports on accurate simulations of the t-J model in 1D and 2D lattices by means of NQS based on a recurrent neural network (RNN) architecture focusing on the calculation of dispersion relations, for which a general method is introduced, and on the performance of the RNN ansatz upon doping.

Automated learning from data by means of deep neural networks is finding use in an ever-increasing number of applications, yet key theoretical questions about how it works remain unanswered. A physics-based approach may help to bridge this gap.

As artificial intelligence (AI) makes increasingly impressive contributions to science, scientists increasingly want to understand how AI reaches its conclusions. Matthew D. Schwartz discusses what it means to understand AI and whether such a goal is achievable — or even needed.

Scientific understanding is one of the main aims of science. This Perspective discusses how advanced computational systems, and artificial intelligence in particular, can contribute to driving scientific understanding.

The phrase ‘arrow of time’ refers to the asymmetry in the flow of events. A machine learning algorithm trained to infer its direction identifies entropy production as the relevant underlying physical principle in the decision-making process.

The application and development of machine-learning methods used in experiments at the frontiers of particle physics (such as the Large Hadron Collider) are reviewed, including recent advances based on deep learning.

Neuromorphic computing takes inspiration from the brain to create energy-efficient hardware for information processing, capable of highly sophisticated tasks. Including more physics in the algorithms and nanoscale materials used for computing could have a major impact in this field.

While widely adopted, contrastive divergence methods for Restricted Boltzmann Machines typically result in poor representations of the data distribution. Here, the authors propose an unsupervised training where gradient-descent is combined with the Machine’s mode samples, significantly improving the final model quality.

The success of machine learning techniques in handling big data sets proves ideal for classifying condensed-matter phases and phase transitions. The technique is even amenable to detecting non-trivial states lacking in conventional order.

Materials simulations are now ubiquitous for explaining material properties. This Review discusses how machine-learned potentials break the limitations of system-size or accuracy, how active-learning will aid their development, how they are applied, and how they may become a more widely used approach.

Recurrent neural networks (RNNs) can learn to process temporal information, such as speech or movement. New work makes such approaches more powerful and flexible by describing theory and experiments demonstrating that RNNs can learn from a few examples to generalize and predict complex dynamics including chaotic behaviour.

Particle tracking velocimetry to estimate particle displacements in fluid flows in complex experimental scenarios is a challenging task and often comes with high computational cost. Liang and colleagues propose a graph neural network and optimal transport-based algorithm that can greatly improve the accuracy of existing tracking algorithms in real-world applications.

A memristor-based annealing system that uses an analogue neuromorphic architecture based on a Hopfield neural network can solve non-deterministic polynomial (NP)-hard max-cut problems in an approach that is potentially more efficient than current quantum, optical and digital approaches.

Tree-based machine learning models are widely used in domains such as healthcare, finance and public services. The authors present an explanation method for trees that enables the computation of optimal local explanations for individual predictions, and demonstrate their method on three medical datasets.

Machine learning methods are becoming increasingly important in the analysis of large-scale genomic, epigenomic, proteomic and metabolic data sets. In this Review, the authors consider the applications of supervised, semi-supervised and unsupervised machine learning methods to genetic and genomic studies. They provide general guidelines for the selection and application of algorithms that are best suited to particular study designs.

Artificial intelligence (AI) and machine learning (ML) are reshaping antibiotic discovery. In this Review, ML approaches that have been and can be used to address issues hindering antimicrobial peptide identification and development are surveyed.

Deep learning-based in silico retrosynthesis prediction methods are able to accelerate retrosynthetic planning, however, their prediction performance remains limited. Here, the authors develop a semi-template-based deep generative model, G²Retro, that can better predict the reactants for one-step retrosyntheses.

In this Opinion article, Hosny et al. discuss the application of artificial intelligence to image-based tasks in the field of radiology and consider the advantages and challenges of its clinical implementation.

There are contrasting views on how to produce the accurate predictions that are needed to guide climate change adaptation. Here, we argue for harnessing artificial intelligence, building on domain-specific knowledge and generating ensembles of moderately high-resolution (10–50 km) climate simulations as anchors for detailed hazard models.

Piecewise linear neural networks (PWLNNs) are a powerful modelling method, particularly in deep learning. In this Primer, Tao et al. introduce the methodology and theoretical analysis of PWLNNs and some of their applications.

Artificial intelligence (AI) is advancing rapidly and is already starting to transform cancer research and care. Here, the authors outline how AI could be incorporated into liver cancer management, highlighting areas with academic, commercial and clinical potential, as well as ongoing progress and pitfalls.

Machine learning has been applied to numerous stages in the drug discovery pipeline. Here, Vamathevan and colleagues discuss the most useful techniques and how machine learning can promote data-driven decision making in drug discovery and development. They highlight major hurdles in the field, such as the required data characteristics for applying machine learning, which will need to be solved as machine learning matures.

In this Review, the authors describe the latest developments in the use of machine learning to interrogate neurodegenerative disease-related datasets. They discuss applications of machine learning to diagnosis, prognosis and therapeutic development, and the challenges involved in analysing health-care data.

An integrated workflow that uses robotics and machine learning to discover all-natural plastic substitutes with programmable properties is presented.

Machine learning is poised to accelerate the development of technologies for a renewable energy future. This Perspective highlights recent advances and in particular proposes Acc(X)eleration Performance Indicators (XPIs) to measure the effectiveness of platforms developed for accelerated energy materials discovery.

AlphaFold predicts protein structures with an accuracy competitive with experimental structures in the majority of cases using a novel deep learning architecture.

Statistical tools based on machine learning are becoming integrated into chemistry research workflows. We discuss the elements necessary to train reliable, repeatable and reproducible models, and recommend a set of guidelines for machine learning reports.

This Perspective discusses how machine learning can help in studying rare diseases using various emerging approaches.

Here, the authors constructed a deep-learning approach to design closed repeat proteins with central binding pockets—a step towards designing proteins to specifically bind small molecules.

This protocol describes how to use an open-source toolbox, DeepLabCut, to train a deep neural network to precisely track user-defined features with limited training data. This allows noninvasive behavioral tracking of movement.

Autoencoders are versatile tools for molecular informatics with the opportunity for advancing molecule and drug design. In this Review, the authors highlight the active areas of development in the field and explore the challenges that need to be addressed moving forward.

The determination of state variables to describe physical systems is a challenging task. A data-driven approach is proposed to automatically identify state variables for unknown systems from high-dimensional observational data.

A machine learning interatomic potential model is designed and trained on diverse crystal data comprising 89 elements, enabling materials discovery across a vast chemical space without retraining.

Automation and real-time reaction monitoring have enabled data-rich experimentation, which is critically important in navigating the complexities of chemical synthesis. Linking real-time analysis with machine learning and artificial intelligence tools provides the opportunity to accelerate the identification of optimal reaction conditions and facilitate error-free autonomous synthesis. This Comment provides a viewpoint underscoring the growing significance of data-rich experiments and interdisciplinary approaches in driving future progress in synthetic chemistry.