The human brain is a funny thing. Certain memories can stick with us forever: the birth of a child, a car crash, an election day. But we only store some details--the color of the hospital delivery room or the smell of the polling station--while others fade, such as the face of the nurse when that child was born, or what we were wearing during that accident. For Google CEO Sundar Pichai, the day he watched AI rise out of a lab is one he'll remember forever. "This was 2012, in a room with a small team, and there were just a few of us," he tells me. An engineer named Jeff Dean, a legendary programmer at Google who helped build its search engine, had been working on a new project and wanted Pichai to have a look. "Anytime Jeff wants to update you on something, you just get excited by it," he says. Pichai doesn't recall exactly which building he was in when Dean presented his work, though odd details of that day have stuck with him. He remembers standing, rather than sitting, and someone joking about an HR snafu that had designated the newly hired Geoffrey Hinton--the "Father of Deep Learning," an AI researcher for four decades, and, later, a Turing Award winner--as an intern. The future CEO of Google was an SVP at the time, running Chrome and Apps, and he hadn't been thinking about AI.

Google leads the world in research on machine intelligence. Its DeepMind subsidiary, in particular, has an impressive list of achievements under its belt. DeepMind's neural networks have achieved superhuman performance in a wide range of games. These include Atari video games such as Pong, Breakout, and Space Invaders and more complex challenges such as the online multiplayer game Starcraft.

September 19, 2019 - DeepMind, a UK-based company, is teaming up with Google Health to tap into expertise in areas like artificial intelligence, app development, data security, and cloud storage. For the last three years, DeepMind has built a team to tackle complex problems in healthcare. Alongside teams at Google, DeepMind will work to build products that support care teams and improve patient...

Linking facts across documents is a challenging task, as the language used to express the same information in a sentence can vary significantly, which complicates the task of multi-document summarization. Consequently, existing approaches heavily rely on hand-crafted features, which are domain-dependent and hard to craft, or additional annotated data, which is costly to gather. To overcome these limitations, we present a novel method, which makes use of two types of sentence embeddings: universal embeddings, which are trained on a large unrelated corpus, and domain-specific embeddings, which are learned during training. To this end, we develop SemSentSum, a fully data-driven model able to leverage both types of sentence embeddings by building a sentence semantic relation graph. SemSentSum achieves competitive results on two types of summary, consisting of 665 bytes and 100 words. Unlike other state-of-the-art models, neither hand-crafted features nor additional annotated data are necessary, and the method is easily adaptable for other tasks. To our knowledge, we are the first to use multiple sentence embeddings for the task of multi-document summarization.

Many attempts have been done to extend the great success of convolutional neural networks (CNNs) achieved on high-end GPU servers to portable devices such as smart phones. Providing compression and acceleration service of deep learning models on the cloud is therefore of significance and is attractive for end users. However, existing network compression and acceleration approaches usually fine-tuning the svelte model by requesting the entire original training data (\eg ImageNet), which could be more cumbersome than the network itself and cannot be easily uploaded to the cloud. In this paper, we present a novel positive-unlabeled (PU) setting for addressing this problem. In practice, only a small portion of the original training set is required as positive examples and more useful training examples can be obtained from the massive unlabeled data on the cloud through a PU classifier with an attention based multi-scale feature extractor. We further introduce a robust knowledge distillation (RKD) scheme to deal with the class imbalance problem of these newly augmented training examples. The superiority of the proposed method is verified through experiments conducted on the benchmark models and datasets. We can use only $8\%$ of uniformly selected data from the ImageNet to obtain an efficient model with comparable performance to the baseline ResNet-34.

The learning rate is one of the most important hyper-parameters for model training and generalization. However, current hand-designed parametric learning rate schedules offer limited flexibility and the predefined schedule may not match the training dynamics of high dimensional and non-convex optimization problems. In this paper, we propose a reinforcement learning based framework that can automatically learn an adaptive learning rate schedule by leveraging the information from past training histories. The learning rate dynamically changes based on the current training dynamics. To validate this framework, we conduct experiments with different neural network architectures on the Fashion MINIST and CIFAR10 datasets. Experimental results show that the auto-learned learning rate controller can achieve better test results. In addition, the trained controller network is generalizable -- able to be trained on one data set and transferred to new problems.

The attacker introduces a fixed-dimension rectangle. This rectangle can be placed by the adversary anywhere in the image, and the attacker can furthermore introduce l noise inside the rectangle with an exogenously specified high bound null (for example, null 255, which effectively allows addition of arbitrary adversarial noise). This model bears some similarity to l 0 attacks, but the rectangle imposes a contiguity constraint, which reflects common physical limitations. The model is clearly abstract: in practice, for example, adversarial occlusions need not be rectangular or have fixed dimensions (for example, the eyeglass frame attack is clearly not rectangular), but at the same time cannot usually be arbitrarily superimposed on an image, as they are implemented in the physical environment. Nevertheless, the model reflects some of the most important aspects common to many physical attacks, such as stickers placed on an adversarially chosen portion of the object we wish to identify.

The success of deep (reinforcement) learning systems crucially depends on the correct choice of hyperparameters which are notoriously sensitive and expensive to evaluate. Training these systems typically requires running iterative processes over multiple epochs or episodes. Traditional approaches only consider final performances of a hyperparameter although intermediate information from the learning curve is readily available. In this paper, we present a Bayesian optimization approach which exploits the iterative structure of learning algorithms for efficient hyperparameter tuning. First, we transform each training curve into a numeric score. Second, we selectively augment the data using the auxiliary information from the curve. This augmentation step enables modeling efficiency while preventing the ill-conditioned issue of Gaussian process covariance matrix happened when adding the whole curve. We demonstrate the efficiency of our algorithm by tuning hyperparameters for the training of deep reinforcement learning agents and convolutional neural networks. Our algorithm outperforms all existing baselines in identifying optimal hyperparameters in minimal time.

We introduce a framework to study the transformation of problems with manifold constraints into unconstrained problems through parametrizations in terms of a Euclidean space. We call these parametrizations "trivializations". We prove conditions under which a trivialization is sound in the context of gradient-based optimization and we show how two large families of trivializations have overall favorable properties, but also suffer from a performance issue. We then introduce "dynamic trivializations", which solve this problem, and we show how these form a family of optimization methods that lie between trivializations and Riemannian gradient descent, and combine the benefits of both of them. We then show how to implement these two families of trivializations in practice for different matrix manifolds. To this end, we prove a formula for the gradient of the exponential of matrices, which can be of practical interest on its own. Finally, we show how dynamic trivializations improve the performance of existing methods on standard tasks designed to test long-term memory within neural networks.

We propose a multi-level method to increase the accuracy of machine learning algorithms for approximating observables in scientific computing, particularly those that arise in systems modeled by differential equations. The algorithm relies on judiciously combining a large number of computationally cheap training data on coarse resolutions with a few expensive training samples on fine grid resolutions. Theoretical arguments for lowering the generalization error, based on reducing the variance of the underlying maps, are provided and numerical evidence, indicating significant gains over underlying single-level machine learning algorithms, are presented. Moreover, we also apply the multi-level algorithm in the context of forward uncertainty quantification and observe a considerable speed-up over competing algorithms.