We investigate the potential of stochastic neural networks for learning effective waveform-based acoustic models. The waveform-based setting, inherent to fully end-to-end speech recognition systems, is motivated by several comparative studies of automatic and human speech recognition that associate standard non-adaptive feature extraction techniques with information loss, which can adversely affect robustness. Stochastic neural networks, on the other hand, are a class of models capable of incorporating rich regularization mechanisms into the learning process. We consider a deep convolutional neural network that first decomposes speech into frequency sub-bands via an adaptive parametric convolutional block where filters are specified by cosine modulations of compactly supported windows. The network then employs standard non-parametric 1D convolutions to extract relevant spectro-temporal patterns while gradually compressing the structured high dimensional representation generated by the parametric block. We rely on a probabilistic parametrization of the proposed neural architecture and learn the model using stochastic variational inference. This requires evaluation of an analytically intractable integral defining the Kullback–Leibler divergence term responsible for regularization, for which we propose an effective approximation based on the Gauss–Hermite quadrature. Our empirical results demonstrate a superior performance of the proposed approach over comparable waveform-based baselines and indicate that it could lead to robustness. Moreover, the approach outperforms a recently proposed deep convolutional neural network for learning of robust acoustic models with standard FBANK features.

An H -packing of G is a collection of vertex-disjoint subgraphs of G such that each component is isomorphic to H . An H -packing of G is maximal if it cannot be extended to a larger H -packing of G . In this paper we consider problem of random allocation of a sequential resource into blocks of m consecutive units and show how it can be successfully modeled in terms of maximal P m -packings. We enumerate maximal P m -packings of P n of a given cardinality and determine the asymptotic behavior of the enumerating sequences. We also compute the expected size of m -packings and provide a lower bound on the efficiency of block-allocation.

AIMS The long-term outcomes of biolimus-eluting stents (BESs) with biodegradable polymer as compared with bare-metal stent (BMS) in patients with ST-segment elevation myocardial infarction (STEMI) remain unknown. METHODS AND RESULTS We performed a 5-year clinical follow-up of 1157 patients (BES: N = 575 and BMS: N = 582) included in the randomized COMFORTABLE AMI trial. Serial intracoronary imaging of stented segments using both intravascular ultrasound (IVUS) and optical coherence tomography performed at baseline and 13 months follow-up were analysed in 103 patients. At 5 years, BES reduced the risk of major adverse cardiac events [MACE; hazard ratio (HR) 0.56, 95% confidence interval (CI): 0.39-0.79, P = 0.001], driven by lower risks for target vessel-related reinfarction (HR 0.44, 95% CI: 0.22-0.87, P = 0.02) and ischaemia-driven target lesion revascularization (HR 0.41, 95% CI: 0.25-0.66, P < 0.001). Definite stent thrombosis (ST) was recorded in 2.2% and 3.9% (HR 0.57, 95% CI: 0.28-1.16, P = 0.12) with no differences in rates of very late definite ST (1.3% vs. 1.6%, P = 0.77). Optical coherence tomography showed no difference in the frequency of malapposed stent struts at follow-up (BES 0.08% vs. BMS 0.02%, P = 0.10). Uncovered stent struts were rarely observed but more frequent in BES (2.1% vs. 0.15%, P < 0.001). In the IVUS analysis, there was no positive remodelling in either group (external elastic membrane area change BES: -0.63 mm2, 95% CI: -1.44 to 0.39 vs. BMS -1.11 mm2, 95% CI: -2.27 to 0.04, P = 0.07). CONCLUSION Compared with BMS, the implantation of biodegradable polymer-coated BES resulted in a lower 5-year rate of MACE in patients with STEMI undergoing primary percutaneous coronary intervention. At 13 months, vascular healing in treated culprit lesions was almost complete irrespective of stent type. CLINICAL TRIAL REGISTRATION http://www.clinicaltrials.gov. Unique identifier: NCT00962416.

The aim of the research was to examine the linguistic structure of the written formof communication for deaf children through the use of simple and complex writtensentences, questions asked and answers to questions asked, as well as the extent of theused statements and messages in the form of expression. The study was conductedon a sample of 70 deaf children. As a measurement instrument, letters of deafrespondents were used on topics of their choice. In total, 45 communication contents,written correspondence were analyzed. The content analysis method and the methodof descriptive analysis were used for data processing. The research results haveshown that the questions were not grammatically correct but the respondents understoodtheir essence and adequately responded to them, that although sentenceswere not linguistically properly written, deaf children can use the dialogue in writtencommunication, that the respondents have problems in writing complex sentences butthat there is a possibility of using complex sentences, that the highest percentage ofrespondents did not use complex sentences correctly, but over 17 % of respondentscan use syntaxically correctly written complex sentences in the form of communicationthat deaf children understand the messages in a written form of communication.Deaf children in written communication have communication but not the linguisticcompetence.

When agents interact with a complex environment, they must form and maintain beliefs about the relevant aspects of that environment. We propose a way to efficiently train expressive generative models in complex environments. We show that a predictive algorithm with an expressive generative model can form stable belief-states in visually rich and dynamic 3D environments. More precisely, we show that the learned representation captures the layout of the environment as well as the position and orientation of the agent. Our experiments show that the model substantially improves data-efficiency on a number of reinforcement learning (RL) tasks compared with strong model-free baseline agents. We find that predicting multiple steps into the future (overshooting), in combination with an expressive generative model, is critical for stable representations to emerge. In practice, using expressive generative models in RL is computationally expensive and we propose a scheme to reduce this computational burden, allowing us to build agents that are competitive with model-free baselines.

The human genetic diversity of the Americas has been shaped by several events of gene flow that have continued since the Colonial Era and the Atlantic slave trade. Moreover, multiple waves of migration followed by local admixture occurred in the last two centuries, the impact of which has been largely unexplored. Here we compiled a genome-wide dataset of ∼12,000 individuals from twelve American countries and ∼6,000 individuals from worldwide populations and applied haplotype-based methods to investigate how historical movements from outside the New World affected i) the genetic structure, ii) the admixture profile, iii) the demographic history and iv) sex-biased gene-flow dynamics, of the Americas. We revealed a high degree of complexity underlying the genetic contribution of European and African populations in North and South America, from both geographic and temporal perspectives, identifying previously unreported sources related to Italy, the Middle East and to specific regions of Africa.