The difficulties in matching the latent posterior to the prior, balancing powerful posteriors with computational efficiency, and the reduced flexibility of data likelihoods are the biggest challenges in the advancement of Variational Autoencoders. We show that these issues arise due to struggles in marginal divergence minimization, and explore an alternative to using conditional distributions that is inspired by Generative Adversarial Networks. The class probability estimation that GANs offer for marginal divergence minimization uncovers a family of VAE-GAN hybrids, which offer the promise of addressing these major challenges in variational inference. We systematically explore the solutions available for distribution matching, but show that these hybrid methods do not fulfill this promise, and the trade-off between generation and inference that they give rise to remains an ongoing research topic.

With the unprecedented success of machine learning in solving challenging problems across multiple domains, there is increasing interest in leveraging state-of-the art techniques to applications in health care. The community-wide efforts for creating large-scale benchmark repositories, such as MIMIC-III and Physionet CinC challenge [1], have accelerated machine learning research in health care. Furthermore, with increased adoption of automated systems for disease diagnosis, there is a huge opportunity for building robust data-driven solutions that can alleviate pain-points within clinical workflows. Broadly, careful modeling of health care data requires tackling inherent challenges including multivariate measurements, long-range temporal dependencies, and missing information in order to make precise predictions. Despite the success of hand-engineered features in clinical models, more recently, regularized representation learning techniques, such as sparse and deep learning, have been more effective. A thorough experimental study on UCR time-series datasets revealed that simple deep learning architectures using 1-D Convolutional Neural Networks (CNNs) can easily outperform traditional task-specific models built on hand-engineered features [2]. More recently, Recurrent Neural Networks (RNN) based on Long Short Term Memory (LSTM) units have become the de-facto solution for clinical time-series analysis.

We propose a new model for unsupervised document embedding. Leading existing approaches either require complex inference or use recurrent neural networks (RNN) that are difficult to parallelize. We take a different route and develop a convolutional neural network (CNN) embedding model. Our CNN architecture is fully parallelizable resulting in over 10x speedup in inference time over RNN models. Parallelizable architecture enables to train deeper models where each successive layer has increasingly larger receptive field and models longer range semantic structure within the document. We additionally propose a fully unsupervised learning algorithm to train this model based on stochastic forward prediction. Empirical results on two public benchmarks show that our approach produces comparable to state-of-the-art accuracy at a fraction of computational cost.

Learning to recognize touch gestures: recurrent vs. convolutional features and dynamic sampling Abstract-- We propose a fully automatic method for learning gestures on big touch devices in a potentially multi-user context. The goal is to learn general models capable of adapting to different gestures, user styles and hardware variations (e.g. Based on deep neural networks, our method features a novel dynamic sampling and temporal normalization component, transforming variable length gestures into fixed length representations while preserving finger/surface contact transitions, that is, the topology of the signal. This sequential representation is then processed with a convolutional model capable, unlike recurrent networks, of learning hierarchical representations with different levels of abstraction. To demonstrate the interest of the proposed method, we introduce a new touch gestures dataset with 6591 gestures performed by 27 people, which is, up to our knowledge, the first of its kind: a publicly available multi-touch gesture dataset for interaction. We also tested our method on a standard dataset of symbolic touch gesture recognition, the MMG dataset, outperforming the state of the art and reporting close to perfect performance. I. INTRODUCTION Touch screen technology has been widely integrated into many different devices for about a decade, becoming a major interface with different use cases ranging from smartphones to big touch tables. Starting with simple interactions, such as taps or single touch gestures, we are now using these interfaces to perform more and more complex actions, involving multiple touches and/or multiple users. If simple interactions do not require complicated engineering to perform well, advanced manipulations such as navigating through a 3D modelisation or designing a document in parallel with different users still craves for easier and better interactions. As of today, different methods and frameworks for touch gesture recognition were developed (see for instance [15], [28] and [7] for reviews). These methods define a specific model for the class, and it is up to the user to execute the correct protocol.

Here's a brief roundup of some interesting news from the AI world from the past two weeks, beyond what we've already reported. TL;DR: Deep RL sucks – A Google engineer has published a long, detailed blog post explaining the current frustrations in deep reinforcement learning, and why it doesn't live up to the hype. Reinforcement learning makes good headlines. Teaching agents to play games like Go well enough to beat human experts like Ke Jie fuels the man versus machine narrative. But a closer look at deep reinforcement learning, a method of machine learning used to train computers to complete a specific task, shows the practice is riddled with problems.

Why and how that deep learning works well on different tasks remains a mystery from a theoretical perspective. In this paper we draw a geometric picture of the deep learning system by finding its analogies with two existing geometric structures, the geometry of quantum computations and the geometry of the diffeomorphic template matching. In this framework, we give the geometric structures of different deep learning systems including convolutional neural networks, residual networks, recursive neural networks, recurrent neural networks and the equilibrium prapagation framework. We can also analysis the relationship between the geometrical structures and their performance of different networks in an algorithmic level so that the geometric framework may guide the design of the structures and algorithms of deep learning systems.

Privacy policy · Terms of use This is i2kweb version 3.5.2.

AI is no longer some abstract dream for the future. It is here, now and bringing change across industries. According to the Forrester AI Readiness Study, 40 percent of the 717 businesses surveyed said they were planning to use intelligent recommendation solutions and 43 percent were planning to use AI-enhanced advanced analytics. With breakthroughs coming thick and fast in machine learning, especially deep learning models, the AI advantage is becoming far more compelling and is spanning across a wider set of applications. However, there are still several roadblocks to wide-scale implementation of AI which are important to recognize as the appetite within the industry to integrate the technology continues to grow.

We consider the problem of building highlevel, class-specific feature detectors from only unlabeled data. Authors: Quoc V. Le, Marc'Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg S. Corrado, Jeffrey Dean and Andrew Y. Ng. (2012) Andrew Ng Adam Coates Brody Huval Quoc Le Andrew Maas Andrew Saxe Richard Socher Sameep Tandon Tao Wang Description Machine learning is a very successful technology but applying it today often requires spending substantial effort hand-designing features. This is true for applications in vision, audio and text. To address this, Ng's group and others are working on "deep learning" algorithms, which can automatically learn feature representations (often from unlabeled data) thus avoiding a lot of time-consuming engineering. These algorithms are based on building massive artificial neural networks that were loosely inspired by cortical (brain) computations.

In a matter of months, General Data Protection Regulation (GDPR) will become a law throughout Europe, deeming a complete overhaul in the way artificial intelligence techniques are used in business settings. By May 25, the GDPR will become fully enforceable throughout the European Union, states the EU GDPR timeline. The coming deadline, which will be enforced in the next 100 days, has sparked a debate among the AI research community and tech giants who are now scrambling to meet the EU's data privacy and algorithmic fairness guidelines. Well, for the EU citizens, GDPR has strengthened their rights by ushering in a new era by unifying data protection rules and placing new obligations on tech enterprises on the process of collecting personal user data. The forthcoming regulations have firmly divided Europe into two different camps – a) one that welcomes the need for data privacy and algorithmic fairness in society, b) tech giants who are bristling at the thought of new challenges, such as asking for user consent in simpler terms and tackling the black box problem of AI, which would make eventually make it illegal, with fines imposed to the tune of 4 percent of global turnover, reportedly.