Variational inference relies on flexible approximate posterior distributions. Normalizing flows provide a general recipe to construct flexible variational posteriors. We introduce Sylvester normalizing flows, which can be seen as a generalization of planar flows. Sylvester normalizing flows remove the well-known single-unit bottleneck from planar flows, making a single transformation much more flexible. We compare the performance of Sylvester normalizing flows against planar flows and inverse autoregressive flows and demonstrate that they compare favorably on several datasets.

We present a formulation of deep learning that aims at producing a large margin classifier. The notion of margin, minimum distance to a decision boundary, has served as the foundation of several theoretically profound and empirically successful results for both classification and regression tasks. However, most large margin algorithms are applicable only to shallow models with a preset feature representation; and conventional margin methods for neural networks only enforce margin at the output layer. Such methods are therefore not well suited for deep networks. In this work, we propose a novel loss function to impose a margin on any chosen set of layers of a deep network (including input and hidden layers). Our formulation allows choosing any norm on the metric measuring the margin. We demonstrate that the decision boundary obtained by our loss has nice properties compared to standard classification loss functions. Specifically, we show improved empirical results on the MNIST, CIFAR-10 and ImageNet datasets on multiple tasks: generalization from small training sets, corrupted labels, and robustness against adversarial perturbations. The resulting loss is general and complementary to existing data augmentation (such as random/adversarial input transform) and regularization techniques (such as weight decay, dropout, and batch norm).

Recent efforts on combining deep models with probabilistic graphical models are promising in providing flexible models that are also easy to interpret. We propose a variational message-passing algorithm for variational inference in such models. First, we propose structured inference networks that incorporate the structure of the graphical model in the inference network of variational auto-encoders (V AE). Second, we establish conditions under which such inference networks enable fast amortized inference similar to V AE. Finally, we derive a variational message passing algorithm to perform efficient natural-gradient inference while retaining the efficiency of the amortized inference. By simultaneously enabling structured, amortized, and natural-gradient inference for deep structured models, our method simplifies and generalizes existing methods. To analyze real-world data, machine learning relies on models that can extract useful patterns. Deep Neural Networks (DNNs) are a popular choice for this purpose because they can learn flexible representations. Another popular choice are probabilistic graphical models (PGMs) which can find interpretable structures in the data. Recent work on combining these two types of models hopes to exploit their complimentary strengths and provide powerful models that are also easy to interpret (Johnson et al., 2016; Krishnan et al., 2015; Archer et al., 2015; Fraccaro et al., 2016). To apply such hybrid models to real-world problems, we need efficient algorithms that can extract useful structure from the data. For deep learning, stochastic-gradient methods are the most popular choice, e.g., those based on back-propagation.

The practical impact of deep learning on complex supervised learning problems has been significant, so much so that almost every Artificial Intelligence problem, or at least a portion thereof, has been somehow recast as a deep learning problem. The applications appeal is significant, but this appeal is increasingly challenged by what some call the challenge of explainability, or more generally the more traditional challenge of debuggability: if the outcomes of a deep learning process produce unexpected results (e.g., less than expected performance of a classifier), then there is little available in the way of theories or tools to help investigate the potential causes of such unexpected behavior, especially when this behavior could impact people's lives. We describe a preliminary framework to help address this issue, which we call "deep visual explanation" (DVE). "Deep," because it is the development and performance of deep neural network models that we want to understand. "Visual," because we believe that the most rapid insight into a complex multi-dimensional model is provided by appropriate visualization techniques, and "Explanation," because in the spectrum from instrumentation by inserting print statements to the abductive inference of explanatory hypotheses, we believe that the key to understanding deep learning relies on the identification and exposure of hypotheses about the performance behavior of a learned deep model. In the exposition of our preliminary framework, we use relatively straightforward image classification examples and a variety of choices on initial configuration of a deep model building scenario. By careful but not complicated instrumentation, we expose classification outcomes of deep models using visualization, and also show initial results for one potential application of interpretability.

Speech enhancement model is used to map a noisy speech to a clean speech. In the training stage, an objective function is often adopted to optimize the model parameters. However, in most studies, there is an inconsistency between the model optimization criterion and the evaluation criterion on the enhanced speech. For example, in measuring speech intelligibility, most of the evaluation metric is based on a short-time objective intelligibility (STOI) measure, while the frame based minimum mean square error (MMSE) between estimated and clean speech is widely used in optimizing the model. Due to the inconsistency, there is no guarantee that the trained model can provide optimal performance in applications. In this study, we propose an end-to-end utterance-based speech enhancement framework using fully convolutional neural networks (FCN) to reduce the gap between the model optimization and evaluation criterion. Because of the utterance-based optimization, temporal correlation information of long speech segments, or even at the entire utterance level, can be considered when perception-based objective functions are used for the direct optimization. As an example, we implement the proposed FCN enhancement framework to optimize the STOI measure. Experimental results show that the STOI of test speech is better than conventional MMSE-optimized speech due to the consistency between the training and evaluation target. Moreover, by integrating the STOI in model optimization, the intelligibility of human subjects and automatic speech recognition (ASR) system on the enhanced speech is also substantially improved compared to those generated by the MMSE criterion.

One of the biggest confusions about "Artificial Intelligence" is that it is a very vague term. That's because Artificial Intelligence or AI is a term that was coined way back in 1955 with extreme hubris: We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions, and concepts, solve kinds of problems now reserved for humans, and improve themselves. AI is over half a century old and carries with it too much baggage.

A fresh wave of artificial intelligence rolling through Microsoft's language translation technologies is bringing more accurate speech recognition to more of the world's languages and higher quality machine-powered translations to all 60 languages supported by Microsoft's translation technologies. The advances were announced at Microsoft Tech Summit Sydney in Australia on November 16. "We've got a complex machine, and we're innovating on all fronts," said Olivier Fontana, the director of product strategy for Microsoft Translator, a platform for text and speech translation services. As the wave spreads, he added, these machine translation tools are allowing more people to grow businesses, build relationships and experience different cultures. Microsoft's research labs around the world are also building on top of these technologies to help people learn how to speak new languages, including a language learning application for non-native speakers of Chinese that also was announced at this week's tech summit. The new Microsoft Translator advances build on last year's switch to deep neural network-powered machine translations, which offer more fluent, human-sounding translations than the predecessor technology known as statistical machine translation.

Historically, building a system that can answer natural language questions about any image has been considered a very ambitious goal. So, how many players are in the image? Well, we can count them and see that there are eleven players, since we are smart enough not to count the referee, right? Although as humans we can normally perform this task without major inconveniences, the development of a system with these capabilities has always seemed closer to science fiction than to the current possibilities of Artificial Intelligence (AI). However, with the advent of Deep Learning (DL), we have witnessed enormous research progress in Visual Question Answering (VQA), in such a way that systems capable of answering these questions are emerging with promising results. In this article I will briefly go through some of the current datasets, approaches and evaluation metrics in VQA, and on how this challenging task can be applied to real life use cases.

Does GDPR require Machine Learning algorithms to explain their output? Probably not, but experts disagree and there is enough ambiguity to keep lawyers busy. Starting May 25, the European Union will require algorithms to explain their output, making deep learning illegal.--

Game Studios spend millions of dollars and thousands of development hours designing game graphics in trying to make them look as close to reality as possible. While the graphics have looked amazingly realistic in the last few years, it is still easy to distinguish them from the real world. However, with the massive advancements made in the field of image processing using Deep Neural Networks, is it time we can leverage that to improve the graphics while simultaneously also reducing the efforts required to create them? To find out whether the recent developments in deep learning can help me answer my question, I tried to focus on improving the player faces in FIFA using the (in?)famous deepfakes algorithm. It is a Deep Neural Network that can be trained to learn and generate extremely realistic human faces.