When training neural networks as classifiers, it is common to observe an increase in average test loss while still maintaining or improving the overall classification accuracy on the same dataset. In spite of the ubiquity of this phenomenon, it has not been well studied and is often dismissively attributed to an increase in borderline correct classifications. We present an empirical investigation that shows how this phenomenon is actually a result of the differential manner by which test samples are processed. In essence: test loss does not increase overall, but only for a small minority of samples. Large representational capacities allow losses to decrease for the vast majority of test samples at the cost of extreme increases for others. This effect seems to be mainly caused by increased parameter values relating to the correctly processed sample features. Our findings contribute to the practical understanding of a common behaviour of deep neural networks. We also discuss the implications of this work for network optimisation and generalisation.

Distributed deep learning workloads include throughput-intensive training tasks on the GPU clusters, where the Distributed Stochastic Gradient Descent (SGD) incurs significant communication delays after backward propagation, forces workers to wait for the gradient synchronization via a centralized parameter server or directly in decentralized workers. We present CrossoverScheduler, an algorithm that enables communication cycles of a distributed training application to be filled by other applications through pipelining communication and computation. With CrossoverScheduler, the running performance of distributed training can be significantly improved without sacrificing convergence rate and network accuracy. We achieve so by introducing Crossover Synchronization which allows multiple distributed deep learning applications to time-share the same GPU alternately. The prototype of CrossoverScheduler is built and integrated with Horovod. Experiments on a variety of distributed tasks show that CrossoverScheduler achieves 20% \times speedup for image classification tasks on ImageNet dataset.

Optimizing the hyperparameters and architecture of a neural network is a long yet necessary phase in the development of any new application. This consuming process can benefit from the elaboration of strategies designed to quickly discard low quality configurations and focus on more promising candidates. This work aims at enhancing HyperNOMAD, a library that adapts a direct search derivative-free optimization algorithm to tune both the architecture and the training of a neural network simultaneously, by targeting two keys steps of its execution and exploiting cheap approximations in the form of static surrogates to trigger the early stopping of the evaluation of a configuration and the ranking of pools of candidates. These additions to HyperNOMAD are shown to improve on its resources consumption without harming the quality of the proposed solutions.

Despite the superior performance in modeling complex patterns to address challenging problems, the black-box nature of Deep Learning (DL) methods impose limitations to their application in real-world critical domains. The lack of a smooth manner for enabling human reasoning about the black-box decisions hinder any preventive action to unexpected events, in which may lead to catastrophic consequences. To tackle the unclearness from black-box models, interpretability became a fundamental requirement in DL-based systems, leveraging trust and knowledge by providing ways to understand the model's behavior. Although a current hot topic, further advances are still needed to overcome the existing limitations of the current interpretability methods in unsupervised DL-based models for Anomaly Detection (AD). Autoencoders (AE) are the core of unsupervised DL-based for AD applications, achieving best-in-class performance. However, due to their hybrid aspect to obtain the results (by requiring additional calculations out of network), only agnostic interpretable methods can be applied to AE-based AD. These agnostic methods are computationally expensive to process a large number of parameters. In this paper we present the RXP (Residual eXPlainer), a new interpretability method to deal with the limitations for AE-based AD in large-scale systems. It stands out for its implementation simplicity, low computational cost and deterministic behavior, in which explanations are obtained through the deviation analysis of reconstructed input features. In an experiment using data from a real heavy-haul railway line, the proposed method achieved superior performance compared to SHAP, demonstrating its potential to support decision making in large scale critical systems.

We introduce the forward-backward (FB) representation of the dynamics of a reward-free Markov decision process. It provides explicit near-optimal policies for any reward specified a posteriori. During an unsupervised phase, we use reward-free interactions with the environment to learn two representations via off-the-shelf deep learning methods and temporal difference (TD) learning. In the test phase, a reward representation is estimated either from observations or an explicit reward description (e.g., a target state). The optimal policy for that reward is directly obtained from these representations, with no planning. The unsupervised FB loss is well-principled: if training is perfect, the policies obtained are provably optimal for any reward function. With imperfect training, the sub-optimality is proportional to the unsupervised approximation error. The FB representation learns long-range relationships between states and actions, via a predictive occupancy map, without having to synthesize states as in model-based approaches. This is a step towards learning controllable agents in arbitrary black-box stochastic environments. This approach compares well to goal-oriented RL algorithms on discrete and continuous mazes, pixel-based MsPacman, and the FetchReach virtual robot arm. We also illustrate how the agent can immediately adapt to new tasks beyond goal-oriented RL.

Automating a robotic task, e.g., robotic suturing can be very complex and time-consuming. Learning a task model to autonomously perform the task is invaluable making the technology, robotic surgery, accessible for a wider community. The data of robotic surgery can be easily logged where the collected data can be used to learn task models. This will result in reduced time and cost of robotic surgery in which a surgeon can supervise the robot operation or give high-level commands instead of low-level control of the tools. We present a data-set of needle insertion in soft tissue with two arms where Arm 1 inserts the needle into the tissue and Arm 2 actively manipulate the soft tissue to ensure the desired and actual exit points are the same. This is important in real-surgery because suturing without active manipulation of tissue may yield failure of the suturing as the stitch may not grip enough tissue to resist the force applied for the suturing. We present a needle insertion dataset including 60 successful trials recorded by 3 pair of stereo cameras. Moreover, we present Deep-robot Learning from Demonstrations that predicts the desired state of the robot at the time step after t (which the optimal action taken at t yields) by looking at the video of the past time steps, i.e. n step time history where N is the memory time window, of the task execution. The experimental results illustrate our proposed deep model architecture is outperforming the existing methods. Although the solution is not yet ready to be deployed on a real robot, the results indicate the possibility of future development for real robot deployment.

You're probably familiar with deepfakes, the digitally altered "synthetic media" that's capable of fooling people into seeing or hearing things that never actually happened. Adversarial examples are like deepfakes for image-recognition A.I. systems -- and while they don't look even slightly strange to us, they're capable of befuddling the heck out of machines. Several years ago, researchers at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (CSAIL) found that they could fool even sophisticated image recognition algorithms into confusing objects simply by slightly altering their surface texture. In the researchers' demonstration, they showed that it was possible to get a cutting-edge neural network to look at a 3D-printed turtle and see a rifle instead. Or to gaze upon a baseball and come away with the conclusion that it is an espresso.

Natural language processing (NLP) and deep learning are growing in popularity for their use in ML technologies like self-driving cars and speech recognition software. As more companies begin to implement deep learning components and other machine learning practices, the demand for software developers and data scientists with proficiency in deep learning is skyrocketing. Today, we will introduce you to a popular deep learning project, the Text Generator, to familiarize you with important, industry-standard NLP concepts, including Markov chains. By the end of this article, you'll understand how to build a Text Generator component for search engine systems and know how to implement Markov chains for faster predictive models.

MIT researchers have developed a way to produce holograms almost instantly. They say the deep learning-based method is so efficient that it could run on a smartphone. A new method called tensor holography could enable the creation of holograms for virtual reality, 3D printing, medical imaging, and more -- and it can run on a smartphone. Despite years of hype, virtual reality headsets have yet to topple TV or computer screens as the go-to devices for video viewing. One reason: VR can make users feel sick.

Imagine if your digital marketing tools had the capacity to predict the future. What would you do with that crystal ball? Or providing each user a set of search results that have shown to be the most likely to yield a conversion? Recommending a product through a web campaign that can be most effective to prompt an engagement? This is where artificial intelligence is most effective for digital marketers.