The victory in 1997 of the chess-playing computer Deep Blue in a six-game series against the then world champion Gary Kasparov was seen as a significant milestone in the development of artificial intelligence. An even greater challenge remained — the ancient game of Go. Despite decades of refinement, until recently the strongest computers were still playing Go at the level of human amateurs. Enter AlphaGo. Developed by Google DeepMind, this program uses deep neural networks to mimic expert players, and further improves its performance by learning from games played against itself. AlphaGo has achieved a 99% win rate against the strongest other Go programs, and defeated the reigning European champion Fan Hui 5–0 in a tournament match. This is the first time that a computer program has defeated a human professional player in even games, on a full, 19 x 19 board, in even games with no handicap.
The multidisciplinary nature of machine intelligence
Machine Intelligence is a highly multidisciplinary and active field, combining computer science, robotics and cognitive science, with potentially transformative applications in many areas of science, industry and society. Current research aims to develop AI systems with broad applicability that will safely interact with humans and the physical world. Different concepts and approaches – machine learning, symbolic reasoning, cognitive science, developmental psychology, robot control engineering, human-machine interactions among others – are increasingly brought together for such goals. To mark the impending launch of Nature Machine Intelligence, this collection explores recent developments in the field and its wide impact in other areas.
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.
A reinforcement learning approach allows a suitably equipped glider to navigate thermal plumes autonomously in an open field.
Reconstructing images from data, whether for medical or astronomical purposes, hinges on well-defined steps. The data sensor encodes an intermediate representation of the observed object, which is converted into an image by a mathematical operation known as the inversion of the encoding function. This inversion is often plagued by sensor imperfections and noise, requiring extra technique-specific steps to correct them. Here, Matthew Rosen and colleagues present a more unified framework termed 'automated transform by manifold approximation' (AUTOMAP). AUTOMAP tackles image reconstruction as a supervised learning task, which uses appropriate training data to link the sensor data to the output image. The authors implemented AUTOMAP with a deep neural network and tested its flexibility in learning how to reconstruct images for various magnetic resonance imaging acquisition strategies. AUTOMAP reduced artefacts and improved accuracy in images reconstructed from noisy and undersampled acquisitions. The authors expect their framework to apply to other imaging methods.
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.
For an artificial agent to be considered truly intelligent it needs to excel at a variety of tasks considered challenging for humans. To date, it has only been possible to create individual algorithms able to master a single discipline — for example, IBM's Deep Blue beat the human world champion at chess but was not able to do anything else. Now a team working at Google's DeepMind subsidiary has developed an artificial agent — dubbed a deep Q-network — that learns to play 49 classic Atari 2600 'arcade' games directly from sensory experience, achieving performance on a par with that of an expert human player. By combining reinforcement learning (selecting actions that maximize reward — in this case the game score) with deep learning (multilayered feature extraction from high-dimensional data — in this case the pixels), the game-playing agent takes artificial intelligence a step nearer the goal of systems capable of learning a diversity of challenging tasks from scratch.
Accelerating PALM/STORM microscopy with deep learning allows super-resolution imaging of >1,000 cells in a few hours.
Quantum machine learning software could enable quantum computers to learn complex patterns in data more efficiently than classical computers are able to.
Strong gravitational lenses are very useful in astronomy for studying the high-redshift Universe, but determining the lens model is slow, painstaking work that requires expert knowledge of the physical processes involved. Future surveys are expected to yield tens of thousands of new lenses to be analysed. Yashar Hezaveh and colleagues use deep artificial neural networks to show that, through machine learning, they can determine lens models 10 million times faster than and with accuracy comparable to sophisticated models. The software can be run by non-experts and will therefore enable more researchers to analyse the predicted influx of lensing data.
Neural networks trained on data from about 130,000 aftershocks from around 100 large earthquakes improve predictions of the spatial distribution of aftershocks and suggest physical quantities that may control earthquake triggering.
Embedding a deep-learning model in the known structure of cellular systems yields DCell, a ‘visible’ neural network that can be used to mechanistically interpret genotype–phenotype relationships.
A deep-learning algorithm is developed to provide rapid and accurate diagnosis of clinical 3D head CT-scan images to triage and prioritize urgent neurological events, thus potentially accelerating time to diagnosis and care in clinical settings.
A neural-network technique can exploit the power of machine learning to mine the exponentially large data sets characterizing the state space of condensed-matter systems. Topological transitions and many-body localization are first on the list.
This Review Article examines the development of organic neuromorphic devices, considering the different switching mechanisms used in the devices and the challenges the field faces in delivering neuromorphic computing applications.
DNA-strand-displacement reactions are used to implement a neural network that can distinguish complex and noisy molecular patterns from a set of nine possibilities—an improvement on previous demonstrations that distinguished only four simple patterns.
Multi-layered neural architectures that implement learning require elaborate mechanisms for symmetric backpropagation of errors that are biologically implausible. Here the authors propose a simple resolution to this problem of blame assignment that works even with feedback using random synaptic weights.
To celebrate the centenary of the year of Alan Turing's birth, four scientists and entrepreneurs assess the divide between neuroscience and computing.
Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science
Artificial neural networks are artificial intelligence computing methods which are inspired by biological neural networks. Here the authors propose a method to design neural networks as sparse scale-free networks, which leads to a reduction in computational time required for training and inference.
FreemoVR is a virtual reality system for freely moving animals. The versatile platform is demonstrated in various experiments with Drosophila, zebrafish, and mice.
Leaky integrate-and-fire artificial neurons based on diffusive memristors enable unsupervised weight updates of drift-memristor synapses in an integrated convolutional neural network capable of pattern recognition.
Small-scale actuators and robots designed to travel, potentially while carrying cargo, to locations that are inaccessible for their human overlords, tend to struggle to negotiate complex surfaces or terrains. This is pronounced if the miniscule robot has only one style of locomotion, which cannot handle challenges such as slopes, steps or a change in friction. Metin Sitti and colleagues develop a magnetically controlled cuboidal silicone device that exhibits a range of locomotive styles including rolling, crawling, walking, jumping and swimming, which it can switch between depending on the terrain. The device can move from swimming through liquids to transporting itself on a solid surface without physical intervention and can pick up cargo, transport it whilst rolling and deposit it elsewhere.
Autonomous mobile robots would be extremely useful in remote or hostile environments such as space, deep oceans or disaster areas. An outstanding challenge is to make such robots able to recover after damage. Jean-Baptiste Mouret and colleagues have developed a machine learning algorithm that enables damaged robots to quickly regain their ability to perform tasks. When they sustain damage — such as broken or even missing legs — the robots adopt an intelligent trial-and-error approach, trying out possible behaviours that they calculate to be potentially high-performing. After a handful of such experiments they discover, in less than two minutes, a compensatory behaviour that works in spite of the damage.
A robot instructed by a machine learning algorithm and coupled with real-time spectroscopic systems provides fast and accurate reaction outcome predictions and reactivity assessments, leading to the discovery of new reactions.
Tactile sensors provide robots with the ability to interact with humans and the environment with great accuracy, yet technical challenges remain for electronic-skin systems to reach human-level performance.
Microrobots are envisioned to revolutionize microsurgery and targeted drug delivery. Their design, operation, locomotion and interaction with the environment are inspired by microorganisms. This Review highlights soft, responsive and active materials for the development of (semi-)autonomous microrobots.
Stephen Cave and Kanta Dihal revisit the extraordinary history of cultural responses to automata.
Machine Intelligence and Society
Artificial intelligence is now superior to humans in many fully competitive games, such as Chess, Go, and Poker. Here the authors develop a machine-learning algorithm that can cooperate effectively with humans when cooperation is beneficial but nontrivial, something humans are remarkably good at.
Self-driving cars offer a bright future, but only if the public can overcome the psychological challenges that stand in the way of widespread adoption. We discuss three: ethical dilemmas, overreactions to accidents, and the opacity of the cars’ decision-making algorithms — and propose steps towards addressing them.
As machine learning infiltrates society, scientists are trying to help ward off injustice.
The development of autonomous weapon systems, by removing the human element of warfare, could make war crimes and atrocities a thing of the past. But if these systems are unable to respect the principles of humanitarian law, we might create a super-intelligent predator that is beyond our control.
Define an international doctrine for cyberspace skirmishes before they escalate into conventional warfare, urge Mariarosaria Taddeo and Luciano Floridi.