Abstract: Purpose Twenty percent of patients with follicular lymphoma (FL) experience progression of disease (POD) within 2 years of initial chemoimmunotherapy. We analyzed data from the National LymphoCare Study to identify whether prognostic FL factors are associated with early POD and whether patients with early POD are at high risk for death.

Abstract: We present the first full release of a survey of the 150 MHz radio sky, observed with the Giant Metrewave Radio Telescope (GMRT) between April 2010 and March 2012 as part of the TIFR GMRT Sky Survey (TGSS) project. Aimed at producing a reliable compact source survey, our automated data reduction pipeline efficiently processed more than 2000 h of observations with minimal human interaction. Through application of innovative techniques such as image-based flagging, direction-dependent calibration of ionospheric phase errors, correcting for systematic offsets in antenna pointing, and improving the primary beam model, we created good quality images for over 95 percent of the 5336 pointings. Our data release covers 36 900 deg2 (or 3.6 π steradians) of the sky between −53° and +90° declination (Dec), which is 90 percent of the total sky. The majority of pointing images have a noise level below 5 mJy beam-1 with an approximate resolution of 25′′×25′′ (or 25′′×25′′/ cos(Dec−19°) for pointings south of 19° declination). We have produced a catalog of 0.62 Million radio sources derived from an initial, high reliability source extraction at the seven sigma level. For the bulk of the survey, the measured overall astrometric accuracy is better than two arcseconds in right ascension and declination, while the flux density accuracy is estimated at approximately ten percent. Within the scope of the TGSS alternative data release (TGSS ADR) project, the source catalog, as well as 5336 mosaic images (5°×5°) and an image cutout service, are made publicly available at the CDS as a service to the astronomical community. Next to enabling a wide range of different scientific investigations, we anticipate that these survey products will provide a solid reference for various new low-frequency radio aperture array telescopes (LOFAR, LWA, MWA, SKA-low), and can play an important role in characterizing the epoch-of-reionisation (EoR) foreground. The TGSS ADR project aims at continuously improving the quality of the survey data products. Near-future improvements include replacement of bright source snapshot images with archival targeted observations, using new observations to fill the holes in sky coverage and replace very poor quality observational data, and an improved flux calibration strategy for less severely affected observational data.

Abstract: Objective To systematically review studies quantifying the associations of long term (clinic), mid-term (home), and short term (ambulatory) variability in blood pressure, independent of mean blood pressure, with cardiovascular disease events and mortality. Data sources Medline, Embase, Cinahl, and Web of Science, searched to 15 February 2016 for full text articles in English. Eligibility criteria for study selection Prospective cohort studies or clinical trials in adults, except those in patients receiving haemodialysis, where the condition may directly impact blood pressure variability. Standardised hazard ratios were extracted and, if there was little risk of confounding, combined using random effects meta-analysis in main analyses. Outcomes included all cause and cardiovascular disease mortality and cardiovascular disease events. Measures of variability included standard deviation, coefficient of variation, variation independent of mean, and average real variability, but not night dipping or day-night variation. Results 41 papers representing 19 observational cohort studies and 17 clinical trial cohorts, comprising 46 separate analyses were identified. Long term variability in blood pressure was studied in 24 papers, mid-term in four, and short-term in 15 (two studied both long term and short term variability). Results from 23 analyses were excluded from main analyses owing to high risks of confounding. Increased long term variability in systolic blood pressure was associated with risk of all cause mortality (hazard ratio 1.15, 95% confidence interval 1.09 to 1.22), cardiovascular disease mortality (1.18, 1.09 to 1.28), cardiovascular disease events (1.18, 1.07 to 1.30), coronary heart disease (1.10, 1.04 to 1.16), and stroke (1.15, 1.04 to 1.27). Increased mid-term and short term variability in daytime systolic blood pressure were also associated with all cause mortality (1.15, 1.06 to 1.26 and 1.10, 1.04 to 1.16, respectively). Conclusions Long term variability in blood pressure is associated with cardiovascular and mortality outcomes, over and above the effect of mean blood pressure. Associations are similar in magnitude to those of cholesterol measures with cardiovascular disease. Limited data for mid-term and short term variability showed similar associations. Future work should focus on the clinical implications of assessment of variability in blood pressure and avoid the common confounding pitfalls observed to date. Systematic review registration PROSPERO CRD42014015695.

Abstract: Monolayer van der Waals (vdW) magnets provide an exciting opportunity for exploring two-dimensional (2D) magnetism for scientific and technological advances, but the intrinsic ferromagnetism has only been observed at low temperatures. Here, we report the observation of room temperature ferromagnetism in manganese selenide (MnSex) films grown by molecular beam epitaxy (MBE). Magnetic and structural characterization provides strong evidence that, in the monolayer limit, the ferromagnetism originates from a vdW manganese diselenide (MnSe2) monolayer, while for thicker films it could originate from a combination of vdW MnSe2 and/or interfacial magnetism of α-MnSe(111). Magnetization measurements of monolayer MnSex films on GaSe and SnSe2 epilayers show ferromagnetic ordering with a large saturation magnetization of ∼4 Bohr magnetons per Mn, which is consistent with the density functional theory calculations predicting ferromagnetism in monolayer 1T-MnSe2. Growing MnSex films on GaSe up to a high thickness (∼4...

Abstract: Importance Administration of hydroxychloroquine with or without azithromycin for the treatment of coronavirus disease 2019 (COVID-19)–associated pneumonia carries increased risk of corrected QT (QTc) prolongation and cardiac arrhythmias. Objective To characterize the risk and degree of QT prolongation in patients with COVID-19 in association with their use of hydroxychloroquine with or without concomitant azithromycin. Design, Setting, and Participants This was a cohort study performed at an academic tertiary care center in Boston, Massachusetts, of patients hospitalized with at least 1 positive COVID-19 nasopharyngeal polymerase chain reaction test result and clinical findings consistent with pneumonia who received at least 1 day of hydroxychloroquine from March 1, 2020, through April 7, 2020. Main Outcomes and Measures Change in QT interval after receiving hydroxychloroquine with or without azithromycin; occurrence of other potential adverse drug events. Results Among 90 patients given hydroxychloroquine, 53 received concomitant azithromycin; 44 (48.9%) were female, and the mean (SD) body mass index was 31.5 (6.6). Hypertension (in 48 patients [53.3%]) and diabetes mellitus (in 26 patients [28.9%]) were the most common comorbid conditions. The overall median (interquartile range) baseline QTc was 455 (430-474) milliseconds (hydroxychloroquine, 473 [454-487] milliseconds vs hydroxychloroquine and azithromycin, 442 [427-461] milliseconds;P Conclusions and Relevance In this cohort study, patients who received hydroxychloroquine for the treatment of pneumonia associated with COVID-19 were at high risk of QTc prolongation, and concurrent treatment with azithromycin was associated with greater changes in QTc. Clinicians should carefully weigh risks and benefits if considering hydroxychloroquine and azithromycin, with close monitoring of QTc and concomitant medication usage.

Abstract: Flexible electrochemical energy storage (FEES) devices have received great attention as a promising power source for the emerging field of flexible and wearable electronic devices. Carbon nanotubes (CNTs) and graphene have many excellent properties that make them ideally suited for use in FEES devices. A brief definition of FEES devices is provided, followed by a detailed overview of various structural models for achieving different FEES devices. The latest research developments on the use of CNTs and graphene in FEES devices are summarized. Finally, future prospects and important research directions in the areas of CNT- and graphene-based flexible electrode synthesis and device integration are discussed.

Abstract: The ability to tune material properties using gating by electric fields is at the heart of modern electronic technology. It is also a driving force behind recent advances in two-dimensional systems, such as the observation of gate electric-field-induced superconductivity and metal-insulator transitions. Here, we describe an ionic field-effect transistor (termed an iFET), in which gate-controlled Li ion intercalation modulates the material properties of layered crystals of 1T-TaS2. The strong charge doping induced by the tunable ion intercalation alters the energetics of various charge-ordered states in 1T-TaS2 and produces a series of phase transitions in thin-flake samples with reduced dimensionality. We find that the charge-density wave states in 1T-TaS2 collapse in the two-dimensional limit at critical thicknesses. Meanwhile, at low temperatures, the ionic gating induces multiple phase transitions from Mott-insulator to metal in 1T-TaS2 thin flakes, with five orders of magnitude modulation in resistance, and superconductivity emerges in a textured charge-density wave state induced by ionic gating. Our method of gate-controlled intercalation opens up possibilities in searching for novel states of matter in the extreme charge-carrier-concentration limit.

Abstract: Exploratory factor analysis (EFA) is a multivariate statistical method that has become a fundamental tool in the development and validation of psychological theories and measurements. However, rese...

Abstract: We present the cross section for the production of a Higgs boson at hadron colliders at next-to-next-to-next-to-leading order (N3LO) in perturbative QCD. The calculation is based on a method to perform a series expansion of the partonic cross section around the threshold limit to an arbitrary order. We perform this expansion to sufficiently high order to obtain the value of the hadronic cross at N3LO in the large top-mass limit. For renormalization and factorization scales equal to half the Higgs boson mass, the N3LO corrections are of the order of +2.2%. The total scale variation at N3LO is 3%, reducing the uncertainty due to missing higher order QCD corrections by a factor of 3.

Abstract: BACKGROUND Since the publication of the European League Against Rheumatism (EULAR) recommendations for the management of large vessel vasculitis (LVV) in 2009, several relevant randomised clinical trials and cohort analyses have been published, which have the potential to change clinical care and therefore supporting the need to update the original recommendations. METHODS Using EULAR standardised operating procedures for EULAR-endorsed recommendations, the EULAR task force undertook a systematic literature review and sought opinion from 20 experts from 13 countries. We modified existing recommendations and created new recommendations. RESULTS Three overarching principles and 10 recommendations were formulated. We recommend that a suspected diagnosis of LVV should be confirmed by imaging or histology. High dose glucocorticoid therapy (40-60 mg/day prednisone-equivalent) should be initiated immediately for induction of remission in active giant cell arteritis (GCA) or Takayasu arteritis (TAK). We recommend adjunctive therapy in selected patients with GCA (refractory or relapsing disease, presence of an increased risk for glucocorticoid-related adverse events or complications) using tocilizumab. Methotrexate may be used as an alternative. Non-biological glucocorticoid-sparing agents should be given in combination with glucocorticoids in all patients with TAK and biological agents may be used in refractory or relapsing patients. We no longer recommend the routine use of antiplatelet or anticoagulant therapy for treatment of LVV unless it is indicated for other reasons. CONCLUSIONS We have updated the recommendations for the management of LVV to facilitate the translation of current scientific evidence and expert opinion into better management and improved outcome of patients in clinical practice.

Abstract: While becoming inextricable to our daily lives, online social media are blamed for increasing mental health problems in younger people. This systematic review synthesized evidence on the influence ...

Abstract: Emerging variants of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) are of clinical concern. In a cohort of 417 persons who had received the second dose of BNT162b2 (Pfizer-BioNTech) or mRNA-1273 (Moderna) vaccine at least 2 weeks previously, we identified 2 women with vaccine breakthrough infection. Despite evidence of vaccine efficacy in both women, symptoms of coronavirus disease 2019 developed, and they tested positive for SARS-CoV-2 by polymerase-chain-reaction testing. Viral sequencing revealed variants of likely clinical importance, including E484K in 1 woman and three mutations (T95I, del142-144, and D614G) in both. These observations indicate a potential risk of illness after successful vaccination and subsequent infection with variant virus, and they provide support for continued efforts to prevent and diagnose infection and to characterize variants in vaccinated persons. (Funded by the National Institutes of Health and others.).

Abstract: We present an approach that exploits hierarchical Recurrent Neural Networks (RNNs) to tackle the video captioning problem, i.e., generating one or multiple sentences to describe a realistic video. Our hierarchical framework contains a sentence generator and a paragraph generator. The sentence generator produces one simple short sentence that describes a specific short video interval. It exploits both temporal-and spatial-attention mechanisms to selectively focus on visual elements during generation. The paragraph generator captures the inter-sentence dependency by taking as input the sentential embedding produced by the sentence generator, combining it with the paragraph history, and outputting the new initial state for the sentence generator. We evaluate our approach on two large-scale benchmark datasets: YouTubeClips and TACoS-MultiLevel. The experiments demonstrate that our approach significantly outperforms the current state-of-the-art methods with BLEU@4 scores 0.499 and 0.305 respectively.

Abstract: Deep neural networks (DNNs) are vulnerable to adversarial examples, perturbations to correctly classified examples which can cause the model to misclassify. In the image domain, these perturbations can often be made virtually indistinguishable to human perception, causing humans and state-of-the-art models to disagree. However, in the natural language domain, small perturbations are clearly perceptible, and the replacement of a single word can drastically alter the semantics of the document. Given these challenges, we use a black-box population-based optimization algorithm to generate semantically and syntactically similar adversarial examples that fool well-trained sentiment analysis and textual entailment models with success rates of 97% and 70%, respectively. We additionally demonstrate that 92.3% of the successful sentiment analysis adversarial examples are classified to their original label by 20 human annotators, and that the examples are perceptibly quite similar. Finally, we discuss an attempt to use adversarial training as a defense, but fail to yield improvement, demonstrating the strength and diversity of our adversarial examples. We hope our findings encourage researchers to pursue improving the robustness of DNNs in the natural language domain.

Abstract: Generating a description of an image is called image captioning. Image captioning requires recognizing the important objects, their attributes, and their relationships in an image. It also needs to generate syntactically and semantically correct sentences. Deep-learning-based techniques are capable of handling the complexities and challenges of image captioning. In this survey article, we aim to present a comprehensive review of existing deep-learning-based image captioning techniques. We discuss the foundation of the techniques to analyze their performances, strengths, and limitations. We also discuss the datasets and the evaluation metrics popularly used in deep-learning-based automatic image captioning.

Abstract: This paper presents an industry equilibrium model where firms have a choice to engage in corporate social responsibility (CSR) activities. We model CSR as an investment to increase product differen...

Abstract: IMPORTANCE: The prevalence of cardiometabolic multimorbidity is increasing. OBJECTIVE: To estimate reductions in life expectancy associated with cardiometabolic multimorbidity. DESIGN, SETTING, AND PARTICIPANTS: Age- and sex-adjusted mortality rates and hazard ratios (HRs) were calculated using individual participant data from the Emerging Risk Factors Collaboration (689,300 participants; 91 cohorts; years of baseline surveys: 1960-2007; latest mortality follow-up: April 2013; 128,843 deaths). The HRs from the Emerging Risk Factors Collaboration were compared with those from the UK Biobank (499,808 participants; years of baseline surveys: 2006-2010; latest mortality follow-up: November 2013; 7995 deaths). Cumulative survival was estimated by applying calculated age-specific HRs for mortality to contemporary US age-specific death rates. EXPOSURES: A history of 2 or more of the following: diabetes mellitus, stroke, myocardial infarction (MI). MAIN OUTCOMES AND MEASURES: All-cause mortality and estimated reductions in life expectancy. RESULTS: In participants in the Emerging Risk Factors Collaboration without a history of diabetes, stroke, or MI at baseline (reference group), the all-cause mortality rate adjusted to the age of 60 years was 6.8 per 1000 person-years. Mortality rates per 1000 person-years were 15.6 in participants with a history of diabetes, 16.1 in those with stroke, 16.8 in those with MI, 32.0 in those with both diabetes and MI, 32.5 in those with both diabetes and stroke, 32.8 in those with both stroke and MI, and 59.5 in those with diabetes, stroke, and MI. Compared with the reference group, the HRs for all-cause mortality were 1.9 (95% CI, 1.8-2.0) in participants with a history of diabetes, 2.1 (95% CI, 2.0-2.2) in those with stroke, 2.0 (95% CI, 1.9-2.2) in those with MI, 3.7 (95% CI, 3.3-4.1) in those with both diabetes and MI, 3.8 (95% CI, 3.5-4.2) in those with both diabetes and stroke, 3.5 (95% CI, 3.1-4.0) in those with both stroke and MI, and 6.9 (95% CI, 5.7-8.3) in those with diabetes, stroke, and MI. The HRs from the Emerging Risk Factors Collaboration were similar to those from the more recently recruited UK Biobank. The HRs were little changed after further adjustment for markers of established intermediate pathways (eg, levels of lipids and blood pressure) and lifestyle factors (eg, smoking, diet). At the age of 60 years, a history of any 2 of these conditions was associated with 12 years of reduced life expectancy and a history of all 3 of these conditions was associated with 15 years of reduced life expectancy. CONCLUSIONS AND RELEVANCE: Mortality associated with a history of diabetes, stroke, or MI was similar for each condition. Because any combination of these conditions was associated with multiplicative mortality risk, life expectancy was substantially lower in people with multimorbidity.

Abstract: The atomic thickness of two-dimensional (2D) materials provides a unique opportunity to control material properties and engineer new functionalities by electrostatic doping. Electrostatic doping has been demonstrated to tune the electrical and optical properties of 2D materials in a wide range, as well as to drive the electronic phase transitions. The recent discovery of atomically thin magnetic insulators has opened up the prospect of electrical control of magnetism and new devices with unprecedented performance. Here we demonstrate control of the magnetic properties of monolayer and bilayer CrI3 by electrostatic doping using a dual-gate field-effect device structure. In monolayer CrI3, doping significantly modifies the saturation magnetization, coercive force and Curie temperature, showing strengthened (weakened) magnetic order with hole (electron) doping. Remarkably, in bilayer CrI3 doping drastically changes the interlayer magnetic order, causing a transition from an antiferromagnetic ground state in the pristine form to a ferromagnetic ground state above a critical electron density. The result reveals a strongly doping-dependent interlayer exchange coupling, which enables robust switching of magnetization in bilayer CrI3 by small gate voltages.

Abstract: We present the first comprehensive study of r-process element nucleosynthesis in the ejecta of compact binary mergers (CBMs) and their relic black hole (BH)–torus systems. The evolution of the BH–accretion tori is simulated for seconds with a Newtonian hydrodynamics code including viscosity effects, pseudo-Newtonian gravity for rotating BHs, and an energy-dependent two-moment closure scheme for the transport of electron neutrinos and antineutrinos. The investigated cases are guided by relativistic double neutron star (NS–NS) and NS–BH merger models, producing ∼3–6 M_⊙ BHs with rotation parameters of A_BH ∼ 0.8 and tori of 0.03–0.3 M_⊙. Our nucleosynthesis analysis includes the dynamical (prompt) ejecta expelled during the CBM phase and the neutrino and viscously driven outflows of the relic BH–torus systems. While typically ∼20–25 per cent of the initial accretion-torus mass are lost by viscously driven outflows, neutrino-powered winds contribute at most another ∼1 per cent, but neutrino heating enhances the viscous ejecta significantly. Since BH–torus ejecta possess a wide distribution of electron fractions (0.1–0.6) and entropies, they produce heavy elements from A ∼ 80 up to the actinides, with relative contributions of A ≳ 130 nuclei being subdominant and sensitively dependent on BH and torus masses and the exact treatment of shear viscosity. The combined ejecta of CBM and BH–torus phases can reproduce the solar abundances amazingly well for A ≳ 90. Varying contributions of the torus ejecta might account for observed variations of lighter elements with 40 ≤ Z ≤ 56 relative to heavier ones, and a considerable reduction of the prompt ejecta compared to the torus ejecta, e.g. in highly asymmetric NS–BH mergers, might explain the composition of heavy-element deficient stars.

Abstract: Ensembles of neural networks are known to be much more robust and accurate than individual networks. However, training multiple deep networks for model averaging is computationally expensive. In this paper, we propose a method to obtain the seemingly contradictory goal of ensembling multiple neural networks at no additional training cost. We achieve this goal by training a single neural network, converging to several local minima along its optimization path and saving the model parameters. To obtain repeated rapid convergence, we leverage recent work on cyclic learning rate schedules. The resulting technique, which we refer to as Snapshot Ensembling, is simple, yet surprisingly effective. We show in a series of experiments that our approach is compatible with diverse network architectures and learning tasks. It consistently yields lower error rates than state-of-the-art single models at no additional training cost, and compares favorably with traditional network ensembles. On CIFAR-10 and CIFAR-100 our DenseNet Snapshot Ensembles obtain error rates of 3.4% and 17.4% respectively.

Abstract: EFSA was asked to deliver a scientific opinion on acrylamide (AA) in food. AA has widespread uses as an industrial chemical. It is also formed when certain foods are prepared at temperatures above 120 °C and low moisture, especially in foods containing asparagine and reducing sugars. The CONTAM Panel evaluated 43 419 analytical results from food commodities. AA was found at the highest levels in solid coffee substitutes and coffee, and in potato fried products. Mean and 95th percentile dietary AA exposures across surveys and age groups were estimated at 0.4 to 1.9 μg/kg body weight (b.w.) per day and 0.6 to 3.4 μg/kg b.w. per day, respectively. The main contributor to total dietary exposure was generally the category ‘Potato fried products (except potato crisps and snacks)’. Preferences in home-cooking can have a substantial impact on human dietary AA exposure. Upon oral intake, AA is absorbed from the gastrointestinal tract and distributed to all organs. AA is extensively metabolised, mostly by conjugation with glutathione but also by epoxidation to glycidamide (GA). Formation of GA is considered to represent the route underlying the genotoxicity and carcinogenicity of AA. Neurotoxicity, adverse effects on male reproduction, developmental toxicity and carcinogenicity were identified as possible critical endpoints for AA toxicity from experimental animal studies. The data from human studies were inadequate for dose-response assessment. The CONTAM Panel selected BMDL10 values of 0.43 mg/kg b.w. per day for peripheral neuropathy in rats and of 0.17 mg/kg b.w. per day for neoplastic effects in mice. The Panel concluded that the current levels of dietary exposure to AA are not of concern with respect to non-neoplastic effects. However, although the epidemiological associations have not demonstrated AA to be a human carcinogen, the margins of exposure (MOEs) indicate a concern for neoplastic effects based on animal evidence. © European Food Safety Authority, 2015

Abstract: The Concise Guide to PHARMACOLOGY 2017/18 provides concise overviews of the key properties of nearly 1800 human drug targets with an emphasis on selective pharmacology (where available), plus links to an open access knowledgebase of drug targets and their ligands (www.guidetopharmacology.org), which provides more detailed views of target and ligand properties. Although the Concise Guide represents approximately 400 pages, the material presented is substantially reduced compared to information and links presented on the website. It provides a permanent, citable, point-in-time record that will survive database updates. The full contents of this section can be found at http://onlinelibrary.wiley.com/doi/10.1111/bph.13878/full. G protein-coupled receptors are one of the eight major pharmacological targets into which the Guide is divided, with the others being: ligand-gated ion channels, voltage-gated ion channels, other ion channels, nuclear hormone receptors, catalytic receptors, enzymes and transporters. These are presented with nomenclature guidance and summary information on the best available pharmacological tools, alongside key references and suggestions for further reading. The landscape format of the Concise Guide is designed to facilitate comparison of related targets from material contemporary to mid-2017, and supersedes data presented in the 2015/16 and 2013/14 Concise Guides and previous Guides to Receptors and Channels. It is produced in close conjunction with the Nomenclature Committee of the Union of Basic and Clinical Pharmacology (NC-IUPHAR), therefore, providing official IUPHAR classification and nomenclature for human drug targets, where appropriate.

Abstract: Language model pre-training has been shown to capture a surprising amount of world knowledge, crucial for NLP tasks such as question answering. However, this knowledge is stored implicitly in the parameters of a neural network, requiring ever-larger networks to cover more facts. To capture knowledge in a more modular and interpretable way, we augment language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend over documents from a large corpus such as Wikipedia, used during pre-training, fine-tuning and inference. For the first time, we show how to pre-train such a knowledge retriever in an unsupervised manner, using masked language modeling as the learning signal and backpropagating through a retrieval step that considers millions of documents. We demonstrate the effectiveness of Retrieval-Augmented Language Model pre-training (REALM) by fine-tuning on the challenging task of Open-domain Question Answering (Open-QA). We compare against state-of-the-art models for both explicit and implicit knowledge storage on three popular Open-QA benchmarks, and find that we outperform all previous methods by a significant margin (4-16% absolute accuracy), while also providing qualitative benefits such as interpretability and modularity.

Abstract: Neural architecture search (NAS) has been proposed to automatically tune deep neural networks, but existing search algorithms, e.g., NASNet, PNAS, usually suffer from expensive computational cost. Network morphism, which keeps the functionality of a neural network while changing its neural architecture, could be helpful for NAS by enabling more efficient training during the search. In this paper, we propose a novel framework enabling Bayesian optimization to guide the network morphism for efficient neural architecture search. The framework develops a neural network kernel and a tree-structured acquisition function optimization algorithm to efficiently explores the search space. Extensive experiments on real-world benchmark datasets have been done to demonstrate the superior performance of the developed framework over the state-of-the-art methods. Moreover, we build an open-source AutoML system based on our method, namely Auto-Keras. The code and documentation are available at https://autokeras.com. The system runs in parallel on CPU and GPU, with an adaptive search strategy for different GPU memory limits.