PARIS – From family and friends to strangers on the subway and public figures on 24-hour news cycles, humans recognize an astonishing 5,000 faces, scientists said Wednesday in the first study of its kind. Through most of history humans lived in small groups of a hundred or so individuals, a pattern that has changed drastically in recent centuries. A study by scientists at Britain's University of York found that our facial recognition abilities allow us to process the thousands of faces we encounter in busy social environments, on our smartphones and our television screens every day. "In everyday life, we are used to identifying friends, colleagues, and celebrities, and many other people by their faces," Rob Jenkins, from York's Department of Psychology, told AFP. "But no one has established how many faces people actually know." For the study, published in the journal Proceedings of the Royal Society B, Jenkins and his team asked participants to write down as many faces they could remember from their personal lives.

A relatively small group of savvy executives have strategies in place to harness new business process automation technologies and thereby advance their digital transformation agenda. Meanwhile, a much larger group is closely following the market leaders, to explore their lessons-learned from pilot projects. According to the latest worldwide market study by 451 Research, new survey results suggest most organisations are adopting or considering artificial intelligence (AI) and machine learning (ML) due to its commercial growth benefits, rather than the potential to cut jobs. Almost 50 percent of their survey respondents have deployed or plan to deploy machine learning in their organisations within the next 12 months. Therefore, this paints a more optimistic picture of machine learning adoption than is often portrayed by other industry analysts. "Out of many possible benefits we presented to our survey respondents, 49 percent cited gaining competitive advantage as the most significant benefit they have received from the technology," said Nick Patience, vice president at 451 Research.

I'm guessing you are scoffing in disbelief at the very suggestion of this article, but bear with me. A growing number of tech analysts are predicting that in less than 20 years we'll all have stopped owning cars, and, what's more, the internal combustion engine will have been consigned to the dustbin of history. Yes, it's a big claim and you are right to be sceptical, but the argument that a unique convergence of new technology is poised to revolutionise personal transportation is more persuasive than you might think. The central idea is pretty simple: Self-driving electric vehicles organised into an Uber-style network will be able to offer such cheap transport that you'll very quickly - we're talking perhaps a decade - decide you don't need a car any more. And if you're thinking this timescale is wildly optimistic, just recall how rapidly cars replaced horses.

Some game-theoretic solution concepts such as the Shapley value and the Banzhaf index have recently gained popularity as measures of node centrality in networks. While this direction of research is promising, the computational problems that surround it are challenging and have largely been left open. To date there are only a few positive results in the literature, which show that some game-theoretic extensions of degree-, closeness- and betweenness-centrality measures are computable in polynomial time, i.e., without the need to enumerate the exponential number of all possible coalitions. In this article, we show that these results can be extended to a much larger class of centrality measures that are based on a family of solution concepts known as semivalues. The family of semivalues includes, among others, the Shapley value and the Banzhaf index. To this end, we present a generic framework for defining game-theoretic network centralities and prove that all centrality measures that can be expressed in this framework are computable in polynomial time. Using our framework, we present a number of new and polynomial-time computable game-theoretic centrality measures.

Thin film processing by means of sputter deposition inherently depends on the interaction of energetic particles with a target surface and the subsequent particle transport. The length and time scales of the underlying physical phenomena span orders of magnitudes. A theoretical description which bridges all time and length scales is not practically possible. Advantage can be taken particularly from the well-separated time scales of the fundamental surface and plasma processes. Initially, surface properties may be calculated from a surface model and stored for a number of representative cases. Subsequently, the surface data may be provided to gas-phase transport simulations via appropriate model interfaces (e.g., analytic expressions or look-up tables) and utilized to define insertion boundary conditions. During run-time evaluation, however, the maintained surface data may prove to be not sufficient. In this case, missing data may be obtained by interpolation (common), extrapolation (inaccurate), or be supplied on-demand by the surface model (computationally inefficient). In this work, a potential alternative is established based on machine learning techniques using artificial neural networks. As a proof of concept, a multilayer perceptron network is trained and verified with sputtered particle distributions obtained from transport of ions in matter based simulations for Ar projectiles bombarding a Ti-Al composite. It is demonstrated that the trained network is able to predict the sputtered particle distributions for unknown, arbitrarily shaped incident ion energy distributions. It is consequently argued that the trained network may be readily used as a machine learning based model interface (e.g., by quasi-continuously sampling the desired sputtered particle distributions from the network), which is sufficiently accurate also in scenarios which have not been previously trained.

Idioms pose problems to almost all Machine Translation systems. This type of language is very frequent in day-to-day language use and cannot be simply ignored. The recent interest in memory augmented models in the field of Language Modelling has aided the systems to achieve good results by bridging long-distance dependencies. In this paper we explore the use of such techniques into a Neural Machine Translation system to help in translation of idiomatic language.

Neural networks have proven their capabilities by outperforming many other approaches on regression or classification tasks on various kinds of data. Other astonishing results have been achieved using neural nets as data generators, especially in settings of generative adversarial networks (GANs). One special application is the field of image domain translations. Here, the goal is to take an image with a certain style (e.g. a photography) and transform it into another one (e.g. a painting). If such a task is performed for unpaired training examples, the corresponding GAN setting is complex, the neural networks are large, and this leads to a high peak memory consumption during, both, training and evaluation phase. This sets a limit to the highest processable image size. We address this issue by the idea of not processing the whole image at once, but to train and evaluate the domain translation on the level of overlapping image subsamples. This new approach not only enables us to translate high-resolution images that otherwise cannot be processed by the neural network at once, but also allows us to work with comparably small neural networks and with limited hardware resources. Additionally, the number of images required for the training process is significantly reduced. We present high-quality results on images with a total resolution of up to over 50 megapixels and emonstrate that our method helps to preserve local image details while it also keeps global consistency.

Designing the architecture for an artificial neural network is a cumbersome task because of the numerous parameters to configure, including activation functions, layer types, and hyper-parameters. With the large number of parameters for most networks nowadays, it is intractable to find a good configuration for a given task by hand. In this paper an Efficient Global Optimization (EGO) algorithm is adapted to automatically optimize and configure convolutional neural network architectures. A configurable neural network architecture based solely on convolutional layers is proposed for the optimization. Without using any knowledge on the target problem and not using any data augmentation techniques, it is shown that on several image classification tasks this approach is able to find competitive network architectures in terms of prediction accuracy, compared to the best hand-crafted ones in literature. In addition, a very small training budget (200 evaluations and 10 epochs in training) is spent on each optimized architectures in contrast to the usual long training time of hand-crafted networks. Moreover, instead of the standard sequential evaluation in EGO, several candidate architectures are proposed and evaluated in parallel, which saves the execution overheads significantly and leads to an efficient automation for deep neural network design.

In this paper, we present a local information theoretic approach to explicitly learn probabilistic clustering of a discrete random variable. Our formulation yields a convex maximization problem for which it is NPhard to find the global optimum. In order to algorithmically solve this optimization problem, we propose two relaxations that are solved via gradient ascent and alternating maximization. Experiments on the MSR Sentence Completion Challenge, MovieLens 100K, and Reuters21578 datasets demonstrate that our approach is competitive with existing techniques and worthy of further investigation. Clustering is one of many important techniques in unsupervised learning that finds structure in unlabeled data.

The paper introduces a non-linear version of the process convolution formalism for building covariance functions for multi-output Gaussian processes. The non-linearity is introduced via Volterra series, one series per each output. We provide closed-form expressions for the mean function and the covariance function of the approximated Gaussian process at the output of the Volterra series. The mean function and covariance function for the joint Gaussian process are derived using formulae for the product moments of Gaussian variables. We compare the performance of the non-linear model against the classical process convolution approach in one synthetic dataset and two real datasets.