Showing papers by "Cornell University published in 2017"

Densely Connected Convolutional Networks

Abstract: Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections—one between each layer and its subsequent layer—our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less memory and computation to achieve high performance. Code and pre-trained models are available at https://github.com/liuzhuang13/DenseNet.

Feature Pyramid Networks for Object Detection

Abstract: Feature pyramids are a basic component in recognition systems for detecting objects at different scales. But pyramid representations have been avoided in recent object detectors that are based on deep convolutional networks, partially because they are slow to compute and memory intensive. In this paper, we exploit the inherent multi-scale, pyramidal hierarchy of deep convolutional networks to construct feature pyramids with marginal extra cost. A top-down architecture with lateral connections is developed for building high-level semantic feature maps at all scales. This architecture, called a Feature Pyramid Network (FPN), shows significant improvement as a generic feature extractor in several applications. Using a basic Faster R-CNN system, our method achieves state-of-the-art single-model results on the COCO detection benchmark without bells and whistles, surpassing all existing single-model entries including those from the COCO 2016 challenge winners. In addition, our method can run at 5 FPS on a GPU and thus is a practical and accurate solution to multi-scale object detection. Code will be made publicly available.

Focal Loss for Dense Object Detection

Abstract: The highest accuracy object detectors to date are based on a two-stage approach popularized by R-CNN, where a classifier is applied to a sparse set of candidate object locations. In contrast, one-stage detectors that are applied over a regular, dense sampling of possible object locations have the potential to be faster and simpler, but have trailed the accuracy of two-stage detectors thus far. In this paper, we investigate why this is the case. We discover that the extreme foreground-background class imbalance encountered during training of dense detectors is the central cause. We propose to address this class imbalance by reshaping the standard cross entropy loss such that it down-weights the loss assigned to well-classified examples. Our novel Focal Loss focuses training on a sparse set of hard examples and prevents the vast number of easy negatives from overwhelming the detector during training. To evaluate the effectiveness of our loss, we design and train a simple dense detector we call RetinaNet. Our results show that when trained with the focal loss, RetinaNet is able to match the speed of previous one-stage detectors while surpassing the accuracy of all existing state-of-the-art two-stage detectors.

Overall Survival with Combined Nivolumab and Ipilimumab in Advanced Melanoma

Abstract: BackgroundNivolumab combined with ipilimumab resulted in longer progression-free survival and a higher objective response rate than ipilimumab alone in a phase 3 trial involving patients with advanced melanoma. We now report 3-year overall survival outcomes in this trial. MethodsWe randomly assigned, in a 1:1:1 ratio, patients with previously untreated advanced melanoma to receive nivolumab at a dose of 1 mg per kilogram of body weight plus ipilimumab at a dose of 3 mg per kilogram every 3 weeks for four doses, followed by nivolumab at a dose of 3 mg per kilogram every 2 weeks; nivolumab at a dose of 3 mg per kilogram every 2 weeks plus placebo; or ipilimumab at a dose of 3 mg per kilogram every 3 weeks for four doses plus placebo, until progression, the occurrence of unacceptable toxic effects, or withdrawal of consent. Randomization was stratified according to programmed death ligand 1 (PD-L1) status, BRAF mutation status, and metastasis stage. The two primary end points were progression-free survival a...

Advanced capabilities for materials modelling with Quantum ESPRESSO.

Abstract: Quantum ESPRESSO is an integrated suite of open-source computer codes for quantum simulations of materials using state-of-the-art electronic-structure techniques, based on density-functional theory, density-functional perturbation theory, and many-body perturbation theory, within the plane-wave pseudopotential and projector-augmented-wave approaches Quantum ESPRESSO owes its popularity to the wide variety of properties and processes it allows to simulate, to its performance on an increasingly broad array of hardware architectures, and to a community of researchers that rely on its capabilities as a core open-source development platform to implement their ideas In this paper we describe recent extensions and improvements, covering new methodologies and property calculators, improved parallelization, code modularization, and extended interoperability both within the distribution and with external software

Advanced capabilities for materials modelling with Quantum ESPRESSO

Abstract: Quantum ESPRESSO is an integrated suite of open-source computer codes for quantum simulations of materials using state-of-the art electronic-structure techniques, based on density-functional theory, density-functional perturbation theory, and many-body perturbation theory, within the plane-wave pseudo-potential and projector-augmented-wave approaches. Quantum ESPRESSO owes its popularity to the wide variety of properties and processes it allows to simulate, to its performance on an increasingly broad array of hardware architectures, and to a community of researchers that rely on its capabilities as a core open-source development platform to implement theirs ideas. In this paper we describe recent extensions and improvements, covering new methodologies and property calculators, improved parallelization, code modularization, and extended interoperability both within the distribution and with external software.

Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization

Abstract: Gatys et al. recently introduced a neural algorithm that renders a content image in the style of another image, achieving so-called style transfer. However, their framework requires a slow iterative optimization process, which limits its practical application. Fast approximations with feed-forward neural networks have been proposed to speed up neural style transfer. Unfortunately, the speed improvement comes at a cost: the network is usually tied to a fixed set of styles and cannot adapt to arbitrary new styles. In this paper, we present a simple yet effective approach that for the first time enables arbitrary style transfer in real-time. At the heart of our method is a novel adaptive instance normalization (AdaIN) layer that aligns the mean and variance of the content features with those of the style features. Our method achieves speed comparable to the fastest existing approach, without the restriction to a pre-defined set of styles. In addition, our approach allows flexible user controls such as content-style trade-off, style interpolation, color & spatial controls, all using a single feed-forward neural network.

Membership Inference Attacks Against Machine Learning Models

Abstract: We quantitatively investigate how machine learning models leak information about the individual data records on which they were trained. We focus on the basic membership inference attack: given a data record and black-box access to a model, determine if the record was in the model's training dataset. To perform membership inference against a target model, we make adversarial use of machine learning and train our own inference model to recognize differences in the target model's predictions on the inputs that it trained on versus the inputs that it did not train on. We empirically evaluate our inference techniques on classification models trained by commercial "machine learning as a service" providers such as Google and Amazon. Using realistic datasets and classification tasks, including a hospital discharge dataset whose membership is sensitive from the privacy perspective, we show that these models can be vulnerable to membership inference attacks. We then investigate the factors that influence this leakage and evaluate mitigation strategies.

Practical Secure Aggregation for Privacy-Preserving Machine Learning

Abstract: We design a novel, communication-efficient, failure-robust protocol for secure aggregation of high-dimensional data. Our protocol allows a server to compute the sum of large, user-held data vectors from mobile devices in a secure manner (i.e. without learning each user's individual contribution), and can be used, for example, in a federated learning setting, to aggregate user-provided model updates for a deep neural network. We prove the security of our protocol in the honest-but-curious and active adversary settings, and show that security is maintained even if an arbitrarily chosen subset of users drop out at any time. We evaluate the efficiency of our protocol and show, by complexity analysis and a concrete implementation, that its runtime and communication overhead remain low even on large data sets and client pools. For 16-bit input values, our protocol offers $1.73 x communication expansion for 210 users and 220-dimensional vectors, and 1.98 x expansion for 214 users and 224-dimensional vectors over sending data in the clear.

On Calibration of Modern Neural Networks

Abstract: Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

On calibration of modern neural networks

Abstract: Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

Immunogenic cell death in cancer and infectious disease

Abstract: Immunogenicity depends on two key factors: antigenicity and adjuvanticity. The presence of exogenous or mutated antigens explains why infected cells and malignant cells can initiate an adaptive immune response provided that the cells also emit adjuvant signals as a consequence of cellular stress and death. Several infectious pathogens have devised strategies to control cell death and limit the emission of danger signals from dying cells, thereby avoiding immune recognition. Similarly, cancer cells often escape immunosurveillance owing to defects in the molecular machinery that underlies the release of endogenous adjuvants. Here, we review current knowledge on the mechanisms that underlie the activation of immune responses against dying cells and their pathophysiological relevance.

Natural climate solutions

Abstract: Better stewardship of land is needed to achieve the Paris Climate Agreement goal of holding warming to below 2 °C; however, confusion persists about the specific set of land stewardship options available and their mitigation potential. To address this, we identify and quantify "natural climate solutions" (NCS): 20 conservation, restoration, and improved land management actions that increase carbon storage and/or avoid greenhouse gas emissions across global forests, wetlands, grasslands, and agricultural lands. We find that the maximum potential of NCS-when constrained by food security, fiber security, and biodiversity conservation-is 23.8 petagrams of CO2 equivalent (PgCO2e) y-1 (95% CI 20.3-37.4). This is ≥30% higher than prior estimates, which did not include the full range of options and safeguards considered here. About half of this maximum (11.3 PgCO2e y-1) represents cost-effective climate mitigation, assuming the social cost of CO2 pollution is ≥100 USD MgCO2e-1 by 2030. Natural climate solutions can provide 37% of cost-effective CO2 mitigation needed through 2030 for a >66% chance of holding warming to below 2 °C. One-third of this cost-effective NCS mitigation can be delivered at or below 10 USD MgCO2-1 Most NCS actions-if effectively implemented-also offer water filtration, flood buffering, soil health, biodiversity habitat, and enhanced climate resilience. Work remains to better constrain uncertainty of NCS mitigation estimates. Nevertheless, existing knowledge reported here provides a robust basis for immediate global action to improve ecosystem stewardship as a major solution to climate change.

Resting-state connectivity biomarkers define neurophysiological subtypes of depression

Abstract: Biomarkers have transformed modern medicine but remain largely elusive in psychiatry, partly because there is a weak correspondence between diagnostic labels and their neurobiological substrates. Like other neuropsychiatric disorders, depression is not a unitary disease, but rather a heterogeneous syndrome that encompasses varied, co-occurring symptoms and divergent responses to treatment. By using functional magnetic resonance imaging (fMRI) in a large multisite sample (n = 1,188), we show here that patients with depression can be subdivided into four neurophysiological subtypes (‘biotypes’) defined by distinct patterns of dysfunctional connectivity in limbic and frontostriatal networks. Clustering patients on this basis enabled the development of diagnostic classifiers (biomarkers) with high (82–93%) sensitivity and specificity for depression subtypes in multisite validation (n = 711) and out-of-sample replication (n = 477) data sets. These biotypes cannot be differentiated solely on the basis of clinical features, but they are associated with differing clinical-symptom profiles. They also predict responsiveness to transcranial magnetic stimulation therapy (n = 154). Our results define novel subtypes of depression that transcend current diagnostic boundaries and may be useful for identifying the individuals who are most likely to benefit from targeted neurostimulation therapies.

Automated Hate Speech Detection and the Problem of Offensive Language

Abstract: A key challenge for automatic hate-speech detection on social media is the separation of hate speech from other instances of offensive language. Lexical detection methods tend to have low precision because they classify all messages containing particular terms as hate speech and previous work using supervised learning has failed to distinguish between the two categories. We used a crowd-sourced hate speech lexicon to collect tweets containing hate speech keywords. We use crowd-sourcing to label a sample of these tweets into three categories: those containing hate speech, only offensive language, and those with neither. We train a multi-class classifier to distinguish between these different categories. Close analysis of the predictions and the errors shows when we can reliably separate hate speech from other offensive language and when this differentiation is more difficult. We find that racist and homophobic tweets are more likely to be classified as hate speech but that sexist tweets are generally classified as offensive. Tweets without explicit hate keywords are also more difficult to classify.

iRECIST: guidelines for response criteria for use in trials testing immunotherapeutics

Abstract: Tumours respond differently to immunotherapies compared with chemotherapeutic drugs, raising questions about the assessment of changes in tumour burden-a mainstay of evaluation of cancer therapeutics that provides key information about objective response and disease progression. A consensus guideline-iRECIST-was developed by the RECIST working group for the use of modified Response Evaluation Criteria in Solid Tumours (RECIST version 1.1) in cancer immunotherapy trials, to ensure consistent design and data collection, facilitate the ongoing collection of trial data, and ultimate validation of the guideline. This guideline describes a standard approach to solid tumour measurements and definitions for objective change in tumour size for use in trials in which an immunotherapy is used. Additionally, it defines the minimum datapoints required from future trials and those currently in development to facilitate the compilation of a data warehouse to use to later validate iRECIST. An unprecedented number of trials have been done, initiated, or are planned to test new immune modulators for cancer therapy using a variety of modified response criteria. This guideline will allow consistent conduct, interpretation, and analysis of trials of immunotherapies.

Diabetic Neuropathy: A Position Statement by the American Diabetes Association

Abstract: Diabetic neuropathies are the most prevalent chronic complications of diabetes. This heterogeneous group of conditions affects different parts of the nervous system and presents with diverse clinical manifestations. The early recognition and appropriate management of neuropathy in the patient with diabetes is important for a number of reasons: 1. Diabetic neuropathy is a diagnosis of exclusion. Nondiabetic neuropathies may be present in patients with diabetes and may be treatable by specific measures. 2. A number of treatment options exist for symptomatic diabetic neuropathy. 3. Up to 50% of diabetic peripheral neuropathies may be asymptomatic. If not recognized and if preventive foot care is not implemented, patients are at risk for injuries to their insensate feet. 4. Recognition and treatment of autonomic neuropathy may improve symptoms, reduce sequelae, and improve quality of life. Among the various forms of diabetic neuropathy, distal symmetric polyneuropathy (DSPN) and diabetic autonomic neuropathies, particularly cardiovascular autonomic neuropathy (CAN), are by far the most studied (1–4). There are several atypical forms of diabetic neuropathy as well (1–4). Patients with prediabetes may also develop neuropathies that are similar to diabetic neuropathies (5–10). Table 1 provides a comprehensive classification scheme for the diabetic neuropathies. View this table: Table 1 Classification for diabetic neuropathies Due to a lack of treatments that target the underlying nerve damage, prevention is the key component of diabetes care. Screening for symptoms and signs of diabetic neuropathy is also critical in clinical practice, as it may detect the earliest stages of neuropathy, enabling early intervention. Although screening for rarer atypical forms of diabetic neuropathy may be warranted, DSPN and autonomic neuropathy are the most common forms encountered in practice. The strongest available evidence regarding treatment pertains to these forms. This Position Statement is based on several recent technical reviews, to which the reader is referred for detailed discussion …

Janus monolayers of transition metal dichalcogenides

Abstract: Structural symmetry-breaking plays a crucial role in determining the electronic band structures of two-dimensional materials. Tremendous efforts have been devoted to breaking the in-plane symmetry of graphene with electric fields on AB-stacked bilayers or stacked van der Waals heterostructures. In contrast, transition metal dichalcogenide monolayers are semiconductors with intrinsic in-plane asymmetry, leading to direct electronic bandgaps, distinctive optical properties and great potential in optoelectronics. Apart from their in-plane inversion asymmetry, an additional degree of freedom allowing spin manipulation can be induced by breaking the out-of-plane mirror symmetry with external electric fields or, as theoretically proposed, with an asymmetric out-of-plane structural configuration. Here, we report a synthetic strategy to grow Janus monolayers of transition metal dichalcogenides breaking the out-of-plane structural symmetry. In particular, based on a MoS2 monolayer, we fully replace the top-layer S with Se atoms. We confirm the Janus structure of MoSSe directly by means of scanning transmission electron microscopy and energy-dependent X-ray photoelectron spectroscopy, and prove the existence of vertical dipoles by second harmonic generation and piezoresponse force microscopy measurements.

Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization

Abstract: Gatys et al. recently introduced a neural algorithm that renders a content image in the style of another image, achieving so-called style transfer. However, their framework requires a slow iterative optimization process, which limits its practical application. Fast approximations with feed-forward neural networks have been proposed to speed up neural style transfer. Unfortunately, the speed improvement comes at a cost: the network is usually tied to a fixed set of styles and cannot adapt to arbitrary new styles. In this paper, we present a simple yet effective approach that for the first time enables arbitrary style transfer in real-time. At the heart of our method is a novel adaptive instance normalization (AdaIN) layer that aligns the mean and variance of the content features with those of the style features. Our method achieves speed comparable to the fastest existing approach, without the restriction to a pre-defined set of styles. In addition, our approach allows flexible user controls such as content-style trade-off, style interpolation, color & spatial controls, all using a single feed-forward neural network.

Inherent Trade-Offs in the Fair Determination of Risk Scores

Abstract: Recent discussion in the public sphere about algorithmic classification has involved tension between competing notions of what it means for a probabilistic classification to be fair to different groups. We formalize three fairness conditions that lie at the heart of these debates, and we prove that except in highly constrained special cases, there is no method that can satisfy these three conditions simultaneously. Moreover, even satisfying all three conditions approximately requires that the data lie in an approximate version of one of the constrained special cases identified by our theorem. These results suggest some of the ways in which key notions of fairness are incompatible with each other, and hence provide a framework for thinking about the trade-offs between them.

T-cell invigoration to tumour burden ratio associated with anti-PD-1 response

Abstract: Despite the success of monotherapies based on blockade of programmed cell death 1 (PD-1) in human melanoma, most patients do not experience durable clinical benefit. Pre-existing T-cell infiltration and/or the presence of PD-L1 in tumours may be used as indicators of clinical response; however, blood-based profiling to understand the mechanisms of PD-1 blockade has not been widely explored. Here we use immune profiling of peripheral blood from patients with stage IV melanoma before and after treatment with the PD-1-targeting antibody pembrolizumab and identify pharmacodynamic changes in circulating exhausted-phenotype CD8 T cells (Tex cells). Most of the patients demonstrated an immunological response to pembrolizumab. Clinical failure in many patients was not solely due to an inability to induce immune reinvigoration, but rather resulted from an imbalance between T-cell reinvigoration and tumour burden. The magnitude of reinvigoration of circulating Tex cells determined in relation to pretreatment tumour burden correlated with clinical response. By focused profiling of a mechanistically relevant circulating T-cell subpopulation calibrated to pretreatment disease burden, we identify a clinically accessible potential on-treatment predictor of response to PD-1 blockade.

The Gender Wage Gap: Extent, Trends, and Explanations

Abstract: Using Panel Study of Income Dynamics (PSID) microdata over the 1980-2010 period, we provide new empirical evidence on the extent of and trends in the gender wage gap, which declined considerably during this time. By 2010, conventional human capital variables taken together explained little of the gender wage gap, while gender differences in occupation and industry continued to be important. Moreover, the gender pay gap declined much more slowly at the top of the wage distribution than at the middle or bottom and by 2010 was noticeably higher at the top. We then survey the literature to identify what has been learned about the explanations for the gap. We conclude that many of the traditional explanations continue to have salience. Although human-capital factors are now relatively unimportant in the aggregate, women's work force interruptions and shorter hours remain significant in high-skilled occupations, possibly due to compensating differentials. Gender differences in occupations and industries, as well as differences in gender roles and the gender division of labor remain important, and research based on experimental evidence strongly suggests that discrimination cannot be discounted. Psychological attributes or noncognitive skills comprise one of the newer explanations for gender differences in outcomes. Our effort to assess the quantitative evidence on the importance of these factors suggests that they account for a small to moderate portion of the gender pay gap, considerably smaller than, say, occupation and industry effects, though they appear to modestly contribute to these differences.

Pre-metastatic niches: organ-specific homes for metastases

Abstract: It is well established that organs of future metastasis are not passive receivers of circulating tumour cells, but are instead selectively and actively modified by the primary tumour before metastatic spread has even occurred. Sowing the 'seeds' of metastasis requires the action of tumour-secreted factors and tumour-shed extracellular vesicles that enable the 'soil' at distant metastatic sites to encourage the outgrowth of incoming cancer cells. In this Review, we summarize the main processes and new mechanisms involved in the formation of the pre-metastatic niche.

Molecular definitions of autophagy and related processes

Abstract: Over the past two decades, the molecular machinery that underlies autophagic responses has been characterized with ever increasing precision in multiple model organisms. Moreover, it has become clear that autophagy and autophagy-related processes have profound implications for human pathophysiology. However, considerable confusion persists about the use of appropriate terms to indicate specific types of autophagy and some components of the autophagy machinery, which may have detrimental effects on the expansion of the field. Driven by the overt recognition of such a potential obstacle, a panel of leading experts in the field attempts here to define several autophagy-related terms based on specific biochemical features. The ultimate objective of this collaborative exchange is to formulate recommendations that facilitate the dissemination of knowledge within and outside the field of autophagy research.

DNA exonuclease Trex1 regulates radiotherapy-induced tumour immunogenicity.

Abstract: Trex1 is an exonuclease that degrades cytosolic DNA and has been associated with modulation of interferon responses in autoimmunity and viral infections. Here, the authors show that Trex1 attenuates the immunogenicity of cancer cells treated with high radiation doses by d…