Showing papers by "Université de Montréal published in 2018"

Graph Attention Networks

Abstract: We present graph attention networks (GATs), novel neural network architectures that operate on graph-structured data, leveraging masked self-attentional layers to address the shortcomings of prior methods based on graph convolutions or their approximations. By stacking layers in which nodes are able to attend over their neighborhoods' features, we enable (implicitly) specifying different weights to different nodes in a neighborhood, without requiring any kind of costly matrix operation (such as inversion) or depending on knowing the graph structure upfront. In this way, we address several key challenges of spectral-based graph neural networks simultaneously, and make our model readily applicable to inductive as well as transductive problems. Our GAT models have achieved or matched state-of-the-art results across four established transductive and inductive graph benchmarks: the Cora, Citeseer and Pubmed citation network datasets, as well as a protein-protein interaction dataset (wherein test graphs remain unseen during training).

Tisagenlecleucel in Children and Young Adults with B-Cell Lymphoblastic Leukemia

Abstract: Background In a single-center phase 1–2a study, the anti-CD19 chimeric antigen receptor (CAR) T-cell therapy tisagenlecleucel produced high rates of complete remission and was associated with serious but mainly reversible toxic effects in children and young adults with relapsed or refractory B-cell acute lymphoblastic leukemia (ALL) Methods We conducted a phase 2, single-cohort, 25-center, global study of tisagenlecleucel in pediatric and young adult patients with CD19+ relapsed or refractory B-cell ALL The primary end point was the overall remission rate (the rate of complete remission or complete remission with incomplete hematologic recovery) within 3 months Results For this planned analysis, 75 patients received an infusion of tisagenlecleucel and could be evaluated for efficacy The overall remission rate within 3 months was 81%, with all patients who had a response to treatment found to be negative for minimal residual disease, as assessed by means of flow cytometry The rates of event-f

Generative Adversarial Networks: An Overview

Abstract: Generative adversarial networks (GANs) provide a way to learn deep representations without extensively annotated training data. They achieve this by deriving backpropagation signals through a competitive process involving a pair of networks. The representations that can be learned by GANs may be used in a variety of applications, including image synthesis, semantic image editing, style transfer, image superresolution, and classification. The aim of this review article is to provide an overview of GANs for the signal processing community, drawing on familiar analogies and concepts where possible. In addition to identifying different methods for training and constructing GANs, we also point to remaining challenges in their theory and application.

Analysis of shared heritability in common disorders of the brain

Abstract: Disorders of the brain can exhibit considerable epidemiological comorbidity and often share symptoms, provoking debate about their etiologic overlap. We quantified the genetic sharing of 25 brain disorders from genome-wide association studies of 265,218 patients and 784,643 control participants and assessed their relationship to 17 phenotypes from 1,191,588 individuals. Psychiatric disorders share common variant risk, whereas neurological disorders appear more distinct from one another and from the psychiatric disorders. We also identified significant sharing between disorders and a number of brain phenotypes, including cognitive measures. Further, we conducted simulations to explore how statistical power, diagnostic misclassification, and phenotypic heterogeneity affect genetic correlations. These results highlight the importance of common genetic variation as a risk factor for brain disorders and the value of heritability-based methods in understanding their etiology.

Adjuvant Pembrolizumab versus Placebo in Resected Stage III Melanoma

Abstract: Background The programmed death 1 (PD-1) inhibitor pembrolizumab has been found to prolong progression-free and overall survival among patients with advanced melanoma. We conducted a phase 3 double-blind trial to evaluate pembrolizumab as adjuvant therapy in patients with resected, high-risk stage III melanoma. Methods Patients with completely resected stage III melanoma were randomly assigned (with stratification according to cancer stage and geographic region) to receive 200 mg of pembrolizumab (514 patients) or placebo (505 patients) intravenously every 3 weeks for a total of 18 doses (approximately 1 year) or until disease recurrence or unacceptable toxic effects occurred. Recurrence-free survival in the overall intention-to-treat population and in the subgroup of patients with cancer that was positive for the PD-1 ligand (PD-L1) were the primary end points. Safety was also evaluated. Results At a median follow-up of 15 months, pembrolizumab was associated with significantly longer recurrence...

Learning deep representations by mutual information estimation and maximization

Abstract: This work investigates unsupervised learning of representations by maximizing mutual information between an input and the output of a deep neural network encoder. Importantly, we show that structure matters: incorporating knowledge about locality in the input into the objective can significantly improve a representation’s suitability for downstream tasks. We further control characteristics of the representation by matching to a prior distribution adversarially. Our method, which we call Deep InfoMax (DIM), outperforms a number of popular unsupervised learning methods and compares favorably with fully-supervised learning on several classification tasks in with some standard architectures. DIM opens new avenues for unsupervised learning of representations and is an important step towards flexible formulations of representation learning objectives for specific end-goals.

Learning deep representations by mutual information estimation and maximization

Abstract: In this work, we perform unsupervised learning of representations by maximizing mutual information between an input and the output of a deep neural network encoder. Importantly, we show that structure matters: incorporating knowledge about locality of the input to the objective can greatly influence a representation's suitability for downstream tasks. We further control characteristics of the representation by matching to a prior distribution adversarially. Our method, which we call Deep InfoMax (DIM), outperforms a number of popular unsupervised learning methods and competes with fully-supervised learning on several classification tasks. DIM opens new avenues for unsupervised learning of representations and is an important step towards flexible formulations of representation-learning objectives for specific end-goals.

Apalutamide Treatment and Metastasis-free Survival in Prostate Cancer

Abstract: Background Apalutamide, a competitive inhibitor of the androgen receptor, is under development for the treatment of prostate cancer. We evaluated the efficacy of apalutamide in men with nonmetastatic castration-resistant prostate cancer who were at high risk for the development of metastasis. Methods We conducted a double-blind, placebo-controlled, phase 3 trial involving men with nonmetastatic castration-resistant prostate cancer and a prostate-specific antigen doubling time of 10 months or less. Patients were randomly assigned, in a 2:1 ratio, to receive apalutamide (240 mg per day) or placebo. All the patients continued to receive androgen-deprivation therapy. The primary end point was metastasis-free survival, which was defined as the time from randomization to the first detection of distant metastasis on imaging or death. Results A total of 1207 men underwent randomization (806 to the apalutamide group and 401 to the placebo group). In the planned primary analysis, which was performed after ...

HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering

Abstract: Existing question answering (QA) datasets fail to train QA systems to perform complex reasoning and provide explanations for answers We introduce HotpotQA, a new dataset with 113k Wikipedia-based question-answer pairs with four key features: (1) the questions require finding and reasoning over multiple supporting documents to answer; (2) the questions are diverse and not constrained to any pre-existing knowledge bases or knowledge schemas; (3) we provide sentence-level supporting facts required for reasoning, allowing QA systems to reason with strong supervision and explain the predictions; (4) we offer a new type of factoid comparison questions to test QA systems’ ability to extract relevant facts and perform necessary comparison We show that HotpotQA is challenging for the latest QA systems, and the supporting facts enable models to improve performance and make explainable predictions

Deep Graph Infomax.

Abstract: We present Deep Graph Infomax (DGI), a general approach for learning node representations within graph-structured data in an unsupervised manner. DGI relies on maximizing mutual information between patch representations and corresponding high-level summaries of graphs---both derived using established graph convolutional network architectures. The learnt patch representations summarize subgraphs centered around nodes of interest, and can thus be reused for downstream node-wise learning tasks. In contrast to most prior approaches to unsupervised learning with GCNs, DGI does not rely on random walk objectives, and is readily applicable to both transductive and inductive learning setups. We demonstrate competitive performance on a variety of node classification benchmarks, which at times even exceeds the performance of supervised learning.

Mutual Information Neural Estimation.

Abstract: We argue that the estimation of mutual information between high dimensional continuous random variables can be achieved by gradient descent over neural networks. We present a Mutual Information Neural Estimator (MINE) that is linearly scalable in dimensionality as well as in sample size, trainable through back-prop, and strongly consistent. We present a handful of applications on which MINE can be used to minimize or maximize mutual information. We apply MINE to improve adversarially trained generative models. We also use MINE to implement the Information Bottleneck, applying it to supervised classification; our results demonstrate substantial improvement in flexibility and performance in these settings.

Personalizing Dialogue Agents: I have a dog, do you have pets too?

Abstract: Chit-chat models are known to have several problems: they lack specificity, do not display a consistent personality and are often not very captivating. In this work we present the task of making chit-chat more engaging by conditioning on profile information. We collect data and train models to (i)condition on their given profile information; and (ii) information about the person they are talking to, resulting in improved dialogues, as measured by next utterance prediction. Since (ii) is initially unknown our model is trained to engage its partner with personal topics, and we show the resulting dialogue can be used to predict profile information about the interlocutors.

Enzalutamide in men with nonmetastatic, castration-resistant prostate cancer

Abstract: Background Men with nonmetastatic, castration-resistant prostate cancer and a rapidly rising prostate-specific antigen (PSA) level are at high risk for metastasis. We hypothesized that enz...

FiLM: Visual Reasoning with a General Conditioning Layer

Abstract: We introduce a general-purpose conditioning method for neu-ral networks called FiLM: Feature-wise Linear Modulation. FiLM layers influence neural network computation via a simple , feature-wise affine transformation based on conditioning information. We show that FiLM layers are highly effective for visual reasoning — answering image-related questions which require a multi-step, high-level process — a task which has proven difficult for standard deep learning methods that do not explicitly model reasoning. Specifically, we show on visual reasoning tasks that FiLM layers 1) halve state-of-the-art error for the CLEVR benchmark, 2) modulate features in a coherent manner, 3) are robust to ablations and architectural modifications, and 4) generalize well to challenging, new data from few examples or even zero-shot.

Maintenance, reserve and compensation: the cognitive neuroscience of healthy ageing

Abstract: Cognitive ageing research examines the cognitive abilities that are preserved and/or those that decline with advanced age. There is great individual variability in cognitive ageing trajectories. Some older adults show little decline in cognitive ability compared with young adults and are thus termed ‘optimally ageing’. By contrast, others exhibit substantial cognitive decline and may develop dementia. Human neuroimaging research has led to a number of important advances in our understanding of the neural mechanisms underlying these two outcomes. However, interpreting the age-related changes and differences in brain structure, activation and functional connectivity that this research reveals is an ongoing challenge. Ambiguous terminology is a major source of difficulty in this venture. Three terms in particular — compensation, maintenance and reserve — have been used in a number of different ways, and researchers continue to disagree about the kinds of evidence or patterns of results that are required to interpret findings related to these concepts. As such inconsistencies can impede progress in both theoretical and empirical research, here, we aim to clarify and propose consensual definitions of these terms. Age-related changes in cognitive ability are the focus of a growing field of research. Cabeza, Rajah and colleagues aim to promote clarity in the field by agreeing upon consensual definitions for three widely discussed concepts: maintenance, compensation and reserve.

Volatile chemical products emerging as largest petrochemical source of urban organic emissions.

Abstract: A gap in emission inventories of urban volatile organic compound (VOC) sources, which contribute to regional ozone and aerosol burdens, has increased as transportation emissions in the United States and Europe have declined rapidly. A detailed mass balance demonstrates that the use of volatile chemical products (VCPs)—including pesticides, coatings, printing inks, adhesives, cleaning agents, and personal care products—now constitutes half of fossil fuel VOC emissions in industrialized cities. The high fraction of VCP emissions is consistent with observed urban outdoor and indoor air measurements. We show that human exposure to carbonaceous aerosols of fossil origin is transitioning away from transportation-related sources and toward VCPs. Existing U.S. regulations on VCPs emphasize mitigating ozone and air toxics, but they currently exempt many chemicals that lead to secondary organic aerosols.

Speaker Recognition from Raw Waveform with SincNet

Abstract: Deep learning is progressively gaining popularity as a viable alternative to i-vectors for speaker recognition. Promising results have been recently obtained with Convolutional Neural Networks (CNNs) when fed by raw speech samples directly. Rather than employing standard hand-crafted features, the latter CNNs learn low-level speech representations from waveforms, potentially allowing the network to better capture important narrow-band speaker characteristics such as pitch and formants. Proper design of the neural network is crucial to achieve this goal.This paper proposes a novel CNN architecture, called SincNet, that encourages the first convolutional layer to discover more meaningful filters. SincNet is based on parametrized sinc functions, which implement band-pass filters. In contrast to standard CNNs, that learn all elements of each filter, only low and high cutoff frequencies are directly learned from data with the proposed method. This offers a very compact and efficient way to derive a customized filter bank specifically tuned for the desired application.Our experiments, conducted on both speaker identification and speaker verification tasks, show that the proposed architecture converges faster and performs better than a standard CNN on raw waveforms.

The state of OA: a large-scale analysis of the prevalence and impact of Open Access articles

Abstract: Despite growing interest in Open Access (OA) to scholarly literature, there is an unmet need for large-scale, up-to-date, and reproducible studies assessing the prevalence and characteristics of OA. We address this need using oaDOI, an open online service that determines OA status for 67 million articles. We use three samples, each of 100,000 articles, to investigate OA in three populations: (1) all journal articles assigned a Crossref DOI, (2) recent journal articles indexed in Web of Science, and (3) articles viewed by users of Unpaywall, an open-source browser extension that lets users find OA articles using oaDOI. We estimate that at least 28% of the scholarly literature is OA (19M in total) and that this proportion is growing, driven particularly by growth in Gold and Hybrid. The most recent year analyzed (2015) also has the highest percentage of OA (45%). Because of this growth, and the fact that readers disproportionately access newer articles, we find that Unpaywall users encounter OA quite frequently: 47% of articles they view are OA. Notably, the most common mechanism for OA is not Gold, Green, or Hybrid OA, but rather an under-discussed category we dub Bronze: articles made free-to-read on the publisher website, without an explicit Open license. We also examine the citation impact of OA articles, corroborating the so-called open-access citation advantage: accounting for age and discipline, OA articles receive 18% more citations than average, an effect driven primarily by Green and Hybrid OA. We encourage further research using the free oaDOI service, as a way to inform OA policy and practice.

HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering

Abstract: Existing question answering (QA) datasets fail to train QA systems to perform complex reasoning and provide explanations for answers. We introduce HotpotQA, a new dataset with 113k Wikipedia-based question-answer pairs with four key features: (1) the questions require finding and reasoning over multiple supporting documents to answer; (2) the questions are diverse and not constrained to any pre-existing knowledge bases or knowledge schemas; (3) we provide sentence-level supporting facts required for reasoning, allowing QA systems to reason with strong supervision and explain the predictions; (4) we offer a new type of factoid comparison questions to test QA systems' ability to extract relevant facts and perform necessary comparison. We show that HotpotQA is challenging for the latest QA systems, and the supporting facts enable models to improve performance and make explainable predictions.

International consensus on the assessment of bruxism : Report of a work in progress

Abstract: In 2013, consensus was obtained on a definition of bruxism as repetitive masticatory muscle activity characterised by clenching or grinding of the teeth and/or by bracing or thrusting of the mandible and specified as either sleep bruxism or awake bruxism. In addition, a grading system was proposed to determine the likelihood that a certain assessment of bruxism actually yields a valid outcome. This study discusses the need for an updated consensus and has the following aims: (i) to further clarify the 2013 definition and to develop separate definitions for sleep and awake bruxism; (ii) to determine whether bruxism is a disorder rather than a behaviour that can be a risk factor for certain clinical conditions; (iii) to re-examine the 2013 grading system; and (iv) to develop a research agenda. It was concluded that: (i) sleep and awake bruxism are masticatory muscle activities that occur during sleep (characterised as rhythmic or non-rhythmic) and wakefulness (characterised by repetitive or sustained tooth contact and/or by bracing or thrusting of the mandible), respectively; (ii) in otherwise healthy individuals, bruxism should not be considered as a disorder, but rather as a behaviour that can be a risk (and/or protective) factor for certain clinical consequences; (iii) both non-instrumental approaches (notably self-report) and instrumental approaches (notably electromyography) can be employed to assess bruxism; and (iv) standard cut-off points for establishing the presence or absence of bruxism should not be used in otherwise healthy individuals; rather, bruxism-related masticatory muscle activities should be assessed in the behaviour's continuum.

Machine Learning for Combinatorial Optimization: a Methodological Tour d'Horizon

Abstract: This paper surveys the recent attempts, both from the machine learning and operations research communities, at leveraging machine learning to solve combinatorial optimization problems. Given the hard nature of these problems, state-of-the-art algorithms rely on handcrafted heuristics for making decisions that are otherwise too expensive to compute or mathematically not well defined. Thus, machine learning looks like a natural candidate to make such decisions in a more principled and optimized way. We advocate for pushing further the integration of machine learning and combinatorial optimization and detail a methodology to do so. A main point of the paper is seeing generic optimization problems as data points and inquiring what is the relevant distribution of problems to use for learning on a given task.

Deep Graph Infomax

Abstract: We present Deep Graph Infomax (DGI), a general approach for learning node representations within graph-structured data in an unsupervised manner. DGI relies on maximizing mutual information between patch representations and corresponding high-level summaries of graphs—both derived using established graph convolutional network architectures. The learnt patch representations summarize subgraphs centered around nodes of interest, and can thus be reused for downstream node-wise learning tasks. In contrast to most prior approaches to unsupervised learning with GCNs, DGI does not rely on random walk objectives, and is readily applicable to both transductive and inductive learning setups. We demonstrate competitive performance on a variety of node classification benchmarks, which at times even exceeds the performance of supervised learning.

Global Matrix 3.0 Physical Activity Report Card Grades for Children and Youth: Results and Analysis From 49 Countries

Abstract: Background: Accumulating sufficient moderate to vigorous physical activity is recognized as a key determinant of physical, physiological, developmental, mental, cognitive, and social health among children and youth (aged 5–17 y). The Global Matrix 3.0ofReportCardgradesonphysicalactivitywasdevelopedtoachieveabetterunderstandingoftheglobalvariationinchildand youth physical activity and associated supports. Methods: Work groups from 49 countries followed harmonized procedures to develop their Report Cards by grading 10 common indicators using the best available data. The participating countries were divided into 3 categories using the United Nations’ human development index (HDI) classification (low or medium, high, and very high HDI). Results: A total of 490 grades, including 369 letter grades and 121 incomplete grades, were assigned by the 49 work groups. Overall, an average grade of “C−,”“D+,” and “C−” was obtained for the low and medium HDI countries, high HDI countries, and very high HDI countries, respectively. Conclusions: The present study provides rich new evidence showing that the situation regarding the physical activity of children and youth is a concern worldwide. Strategic public investments to implement effective interventions to increase physical activity opportunities are needed.

On the Spectral Bias of Neural Networks

Abstract: Neural networks are known to be a class of highly expressive functions able to fit even random input-output mappings with $100\%$ accuracy. In this work, we present properties of neural networks that complement this aspect of expressivity. By using tools from Fourier analysis, we show that deep ReLU networks are biased towards low frequency functions, meaning that they cannot have local fluctuations without affecting their global behavior. Intuitively, this property is in line with the observation that over-parameterized networks find simple patterns that generalize across data samples. We also investigate how the shape of the data manifold affects expressivity by showing evidence that learning high frequencies gets \emph{easier} with increasing manifold complexity, and present a theoretical understanding of this behavior. Finally, we study the robustness of the frequency components with respect to parameter perturbation, to develop the intuition that the parameters must be finely tuned to express high frequency functions.

Sex differences in Alzheimer disease — the gateway to precision medicine

Abstract: Alzheimer disease (AD) is characterized by wide heterogeneity in cognitive and behavioural syndromes, risk factors and pathophysiological mechanisms. Addressing this phenotypic variation will be crucial for the development of precise and effective therapeutics in AD. Sex-related differences in neural anatomy and function are starting to emerge, and sex might constitute an important factor for AD patient stratification and personalized treatment. Although the effects of sex on AD epidemiology are currently the subject of intense investigation, the notion of sex-specific clinicopathological AD phenotypes is largely unexplored. In this Review, we critically discuss the evidence for sex-related differences in AD symptomatology, progression, biomarkers, risk factor profiles and treatment. The cumulative evidence reviewed indicates sex-specific patterns of disease manifestation as well as sex differences in the rates of cognitive decline and brain atrophy, suggesting that sex is a crucial variable in disease heterogeneity. We discuss critical challenges and knowledge gaps in our current understanding. Elucidating sex differences in disease phenotypes will be instrumental in the development of a 'precision medicine' approach in AD, encompassing individual, multimodal, biomarker-driven and sex-sensitive strategies for prevention, detection, drug development and treatment.

Engaging patients to improve quality of care: a systematic review

Abstract: To identify the strategies and contextual factors that enable optimal engagement of patients in the design, delivery, and evaluation of health services. We searched MEDLINE, EMBASE, CINAHL, Cochrane, Scopus, PsychINFO, Social Science Abstracts, EBSCO, and ISI Web of Science from 1990 to 2016 for empirical studies addressing the active participation of patients, caregivers, or families in the design, delivery and evaluation of health services to improve quality of care. Thematic analysis was used to identify (1) strategies and contextual factors that enable optimal engagement of patients, (2) outcomes of patient engagement, and (3) patients’ experiences of being engaged. Forty-eight studies were included. Strategies and contextual factors that enable patient engagement were thematically grouped and related to techniques to enhance design, recruitment, involvement and leadership action, and those aimed to creating a receptive context. Reported outcomes ranged from educational or tool development and informed policy or planning documents (discrete products) to enhanced care processes or service delivery and governance (care process or structural outcomes). The level of engagement appears to influence the outcomes of service redesign—discrete products largely derived from low-level engagement (consultative unidirectional feedback)—whereas care process or structural outcomes mainly derived from high-level engagement (co-design or partnership strategies). A minority of studies formally evaluated patients’ experiences of the engagement process (n = 12; 25%). While most experiences were positive—increased self-esteem, feeling empowered, or independent—some patients sought greater involvement and felt that their involvement was important but tokenistic, especially when their requests were denied or decisions had already been made. Patient engagement can inform patient and provider education and policies, as well as enhance service delivery and governance. Additional evidence is needed to understand patients’ experiences of the engagement process and whether these outcomes translate into improved quality of care. N/A (data extraction completed prior to registration on PROSPERO).