In what was perhaps one of the most memorable openings in literature in English, Charles Dickens began his immortal A Tale of Two Cities with this: "It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity, it was the season of light, it was the season of darkness, it was the spring of hope, it was the winter of despair, we had everything before us, we had nothing before us, we were all going direct o heaven, we were all going direct the other way--in short, the period was so far like the present period, that some of its noisiest authorities insisted on its being received, for good or for evil, in the superlative degree of comparison only." And yes, that was one long, run-on sentence….! And yes, participating in RSNA 2017, this year's edition of the annual RSNA Conference (sponsored by the Oak Brook, Ill.-based Radiological Society of North America), did bring to mind Dickens' astonishing opening to his great 1859 novel. And though I saw no one at RSNA 2017 who reminded me at all of Sydney Carton, Lucie Manette, Charles Darnay, or Madame Defarge, I did actually think a bit about France in 1775 (on the eve of the French Revolution). Here's the thing: the practice of radiology, as we've all known it, is moving into uncharted territory now, as the financial, operational, and medical practice model on which it's been based, is shifting under the feet of today's radiologists. With both Medicare and private-insurer payment under accelerating threat (let's face it, diagnostic imaging procedures are an easy target for reimbursement deficit-hawk types), and with the demands for speed of turnaround for interpretive reports also accelerating, there are literally not enough hours in the day for practicing radiologists to make up growing income shortfalls from ongoing reductions in payment from all sources.

AI With The Best hosted 50 speakers and hundreds of attendees from all over the world on a single platform on October 14-15, 2017. The platform held live talks, Insights/Questions pages, and bookings for 1-on-1s with speakers. Ian Goodfellow is a Research Scientist at Google Brain and presented on Numerical Computation for Deep Learning. This presentation covers chapter 4 of the Deep Learning textbook ( www.deeplearningbook.org). Deep learning algorithms are usually described in terms of real numbers, with infinite precision.

XLearning is a convenient and efficient scheduling platform combined with the big data and artificial intelligence, support for a variety of machine learning, deep learning frameworks. XLearning has the satisfactory scalability and compatibility. Besides the distributed mode of TensorFlow and MXNet frameworks, XLearning supports the standalone mode of all deep learning frameworks such as Caffe, Theano, PyTorch. Moreover, XLearning allows the custom versions and multi-version of frameworks flexibly. XLearning is enable to specify the input strategy for the input data --input by setting the --input-strategy parameter or xlearning.input.strategy

This work was done in collaboration with Ding Ding and Sergey Ermolin from Intel. In recent years, the scale of datasets and models used in deep learning has increased dramatically. Although larger datasets and models can improve the accuracy in many AI applications, they often take much longer to train on a single machine. However, it is not very common to distribute the training to large clusters using current popular deep learning frameworks, compared to what's been long around in the Big Data area, as it's often harder to gain access to a large GPU cluster and lack of convenient facilities in popular DL frameworks for distributed training. By leveraging the cluster distribution capabilities in Apache Spark, BigDL successfully performs very large-scale distributed training and inference.

Almost 15 years after scientists first sequenced the human genome, making sense of the enormous amount of data that encodes human life remains a formidable challenge. But it is also precisely the sort of problem that machine learning excels at. On Monday, Google released a tool called DeepVariant that uses the latest AI techniques to build a more accurate picture of a person's genome from sequencing data. DeepVariant helps turn high-throughput sequencing readouts into a picture of a full genome. It automatically identifies small insertion and deletion mutations and single-base-pair mutations in sequencing data. High-throughput sequencing became widely available in the 2000s and has made genome sequencing more accessible.

In this setting, we only have access to the stochastic function f(x; ξ), where the random variable ξ is sampled from an underlying distribution D. The task is to optimize the expected function f(x), which in general may be nonconvex. This framework covers a wide range of problems, including the offline setting where we minimize the empirical loss over a fixed amount of data, and the online setting where data arrives sequentially. One of the most prominent applications of stochastic optimization has been in large-scale statistics and machine learning problems, such as the optimization of deep neural networks. Classical analysis in nonconvex optimization only guarantees convergence to a first-order stationary point (i.e., a point x satisfying ‖ f(x)‖ 0), which can be a local minimum, a local maximum, or a saddle point. This paper goes further, proposing an algorithm that escapes saddle points and converges to a local minimum.

Location fingerprinting using received signal strengths (RSSs) from wireless network infrastructure is one of the most popular and promising technologies for localization in an indoor environment, where there is no line-of-sight signal from the global positioning system (GPS) available [1]: For example, a vector of pairs of a service set identifier (SSID) and an RSS for a Wi-Fi access point (AP) measured at a location can be its location fingerprint. A location of a user/device then can be estimated by finding the closest match between its RSS measurement and the fingerprints of known locations in a database [2]. Note that the location fingerprinting technique does not require the installation of any new infrastructure or the modification of existing devices, but it is just based on the existing wireless infrastructure, which is its major advantage over alternative techniques. When the indoor localization is to cover a large building complex -- e.g., a big shopping mall or a university campus -- where there are lots of multistory buildings under the same management, the scalability of fingerprinting techniques becomes an important issue. The current state-of-the-art Wi-Fi fingerprinting techniques assume a hierarchical approach to the indoor localization, where the building, floor, and position (e.g., a label or coordinates) of a location are estimated in a hierarchical and sequential way using a different algorithm tailored for each task.

For decades, context-dependent phonemes have been the dominant sub-word unit for conventional acoustic modeling systems. This status quo has begun to be challenged recently by end-to-end models which seek to combine acoustic, pronunciation, and language model components into a single neural network. Such systems, which typically predict graphemes or words, simplify the recognition process since they remove the need for a separate expert-curated pronunciation lexicon to map from phoneme-based units to words. However, there has been little previous work comparing phoneme-based versus grapheme-based sub-word units in the end-to-end modeling framework, to determine whether the gains from such approaches are primarily due to the new probabilistic model, or from the joint learning of the various components with grapheme-based units. In this work, we conduct detailed experiments which are aimed at quantifying the value of phoneme-based pronunciation lexica in the context of end-to-end models. We examine phoneme-based end-to-end models, which are contrasted against grapheme-based ones on a large vocabulary English Voice-search task, where we find that graphemes do indeed outperform phonemes. We also compare grapheme and phoneme-based approaches on a multi-dialect English task, which once again confirm the superiority of graphemes, greatly simplifying the system for recognizing multiple dialects.

Sequence-to-sequence models, such as attention-based models in automatic speech recognition (ASR), are typically trained to optimize the cross-entropy criterion which corresponds to improving the log-likelihood of the data. However, system performance is usually measured in terms of word error rate (WER), not log-likelihood. Traditional ASR systems benefit from discriminative sequence training which optimizes criteria such as the state-level minimum Bayes risk (sMBR) which are more closely related to WER. In the present work, we explore techniques to train attention-based models to directly minimize expected word error rate. We consider two loss functions which approximate the expected number of word errors: either by sampling from the model, or by using N-best lists of decoded hypotheses, which we find to be more effective than the sampling-based method. In experimental evaluations, we find that the proposed training procedure improves performance by up to 8.2% relative to the baseline system. This allows us to train grapheme-based, uni-directional attention-based models which match the performance of a traditional, state-of-the-art, discriminative sequence-trained system on a mobile voice-search task.

ABSTRACT Having a sequence-to-sequence model which can operate in an online fashion is important for streaming applications such as Voice Search. Neural transducer is a streaming sequence-to-sequence model, but has shown a significant degradation in performance compared to nonstreaming models such as Listen, Attend and Spell (LAS). Specifically, we look at increasing the window over which NT computes attention, mainly by looking backwards in time so the model still remains online. In addition, we explore initializing a NT model from a LAS-trained model so that it is guided with a better alignment. Finally, we explore including stronger language models such as using wordpiece models, and applying an external LM during the beam search. On a Voice Search task, we find with these improvements we can get NT to match the performance of LAS. 1. INTRODUCTION Sequence-to-sequence models have become popular in the automatic speech recognition (ASR) community [1, 2, 3, 4], as they allow for one neural network to jointly learn an acoutic, pronunciation and language model, greatly simplifying the ASR pipeline.