Showing papers by "Georgia Institute of Technology published in 2019"

First M87 Event Horizon Telescope Results. I. The Shadow of the Supermassive Black Hole

Abstract: When surrounded by a transparent emission region, black holes are expected to reveal a dark shadow caused by gravitational light bending and photon capture at the event horizon. To image and study this phenomenon, we have assembled the Event Horizon Telescope, a global very long baseline interferometry array observing at a wavelength of 1.3 mm. This allows us to reconstruct event-horizon-scale images of the supermassive black hole candidate in the center of the giant elliptical galaxy M87. We have resolved the central compact radio source as an asymmetric bright emission ring with a diameter of 42 +/- 3 mu as, which is circular and encompasses a central depression in brightness with a flux ratio greater than or similar to 10: 1. The emission ring is recovered using different calibration and imaging schemes, with its diameter and width remaining stable over four different observations carried out in different days. Overall, the observed image is consistent with expectations for the shadow of a Kerr black hole as predicted by general relativity. The asymmetry in brightness in the ring can be explained in terms of relativistic beaming of the emission from a plasma rotating close to the speed of light around a black hole. We compare our images to an extensive library of ray-traced general-relativistic magnetohydrodynamic simulations of black holes and derive a central mass of M = (6.5 +/- 0.7) x 10(9) M-circle dot. Our radio-wave observations thus provide powerful evidence for the presence of supermassive black holes in centers of galaxies and as the central engines of active galactic nuclei. They also present a new tool to explore gravity in its most extreme limit and on a mass scale that was so far not accessible.

Electrospinning and Electrospun Nanofibers: Methods, Materials, and Applications

Abstract: Electrospinning is a versatile and viable technique for generating ultrathin fibers. Remarkable progress has been made with regard to the development of electrospinning methods and engineering of electrospun nanofibers to suit or enable various applications. We aim to provide a comprehensive overview of electrospinning, including the principle, methods, materials, and applications. We begin with a brief introduction to the early history of electrospinning, followed by discussion of its principle and typical apparatus. We then discuss its renaissance over the past two decades as a powerful technology for the production of nanofibers with diversified compositions, structures, and properties. Afterward, we discuss the applications of electrospun nanofibers, including their use as "smart" mats, filtration membranes, catalytic supports, energy harvesting/conversion/storage components, and photonic and electronic devices, as well as biomedical scaffolds. We highlight the most relevant and recent advances related to the applications of electrospun nanofibers by focusing on the most representative examples. We also offer perspectives on the challenges, opportunities, and new directions for future development. At the end, we discuss approaches to the scale-up production of electrospun nanofibers and briefly discuss various types of commercial products based on electrospun nanofibers that have found widespread use in our everyday life.

ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks

Abstract: We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, pro-cessing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.

ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks

Abstract: We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, processing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.

Triboelectric Nanogenerator: A Foundation of the Energy for the New Era

Abstract: The triboelectric effect is ubiquitous in our everyday life and results from two different materials coming into contact. It is generally regarded as a negative effect in industry given that the electrostatic charges induced from it can lead to ignition, dust explosions, dielectric breakdown, electronic damage, etc. From an energy point of view, those electrostatic charges constitute a capacitive energy device when the two triboelectric surfaces are separated, which led to the invention of early electrostatic generators such as the “friction machine” and Van de Graaff generator.[1] As the world is marching into the era of the internet of things (IoTs) and artificial intelligence, the most vital development for hardware is a multifunctional array of sensing systems, which forms the foundation of the fourth industrial revolution toward an intelligent world. Given the need for mobility of these multitudes of sensors, the success of the IoTs calls for distributed energy sources, which can be provided by solar, thermal, wind, and mechanical triggering/vibrations. The triboelectric nanogenerator (TENG) for mechanical energy harvesting developed by Z.L. Wang’s group is one of the best choices for this energy for the new era, since triboelectrification is a universal and ubiquitous effect with an abundant choice of materials. The development of self-powered active sensors enabled by TENGs is revolutionary compared to externally powered passive sensors, similar to the advance from wired to wireless communication. In this paper, the fundamental theory, experiments, and applications of TENGs are reviewed as a foundation of the energy for the new era with four major application fields: micro/nano power sources, self-powered sensors, large-scale blue energy, and direct high-voltage power sources. A roadmap is proposed for the research and commercialization of TENG in the next 10 years.

Scientists' Warning to Humanity: Microorganisms and Climate Change

Abstract: In the Anthropocene, in which we now live, climate change is impacting most life on Earth. Microorganisms support the existence of all higher trophic life forms. To understand how humans and other life forms on Earth (including those we are yet to discover) can withstand anthropogenic climate change, it is vital to incorporate knowledge of the microbial 'unseen majority'. We must learn not just how microorganisms affect climate change (including production and consumption of greenhouse gases) but also how they will be affected by climate change and other human activities. This Consensus Statement documents the central role and global importance of microorganisms in climate change biology. It also puts humanity on notice that the impact of climate change will depend heavily on responses of microorganisms, which are essential for achieving an environmentally sustainable future.

Argoverse: 3D Tracking and Forecasting With Rich Maps

Abstract: We present Argoverse, a dataset designed to support autonomous vehicle perception tasks including 3D tracking and motion forecasting. Argoverse includes sensor data collected by a fleet of autonomous vehicles in Pittsburgh and Miami as well as 3D tracking annotations, 300k extracted interesting vehicle trajectories, and rich semantic maps. The sensor data consists of 360 degree images from 7 cameras with overlapping fields of view, forward-facing stereo imagery, 3D point clouds from long range LiDAR, and 6-DOF pose. Our 290km of mapped lanes contain rich geometric and semantic metadata which are not currently available in any public dataset. All data is released under a Creative Commons license at Argoverse.org. In baseline experiments, we use map information such as lane direction, driveable area, and ground height to improve the accuracy of 3D object tracking. We use 3D object tracking to mine for more than 300k interesting vehicle trajectories to create a trajectory forecasting benchmark. Motion forecasting experiments ranging in complexity from classical methods (k-NN) to LSTMs demonstrate that using detailed vector maps with lane-level information substantially reduces prediction error. Our tracking and forecasting experiments represent only a superficial exploration of the potential of rich maps in robotic perception. We hope that Argoverse will enable the research community to explore these problems in greater depth.

On the Variance of the Adaptive Learning Rate and Beyond

Abstract: The learning rate warmup heuristic achieves remarkable success in stabilizing training, accelerating convergence and improving generalization for adaptive stochastic optimization algorithms like RMSprop and Adam. Here, we study its mechanism in details. Pursuing the theory behind warmup, we identify a problem of the adaptive learning rate (i.e., it has problematically large variance in the early stage), suggest warmup works as a variance reduction technique, and provide both empirical and theoretical evidence to verify our hypothesis. We further propose RAdam, a new variant of Adam, by introducing a term to rectify the variance of the adaptive learning rate. Extensive experimental results on image classification, language modeling, and neural machine translation verify our intuition and demonstrate the effectiveness and robustness of our proposed method. All implementations are available at: this https URL.

Quantifying the triboelectric series.

Abstract: Triboelectrification is a well-known phenomenon that commonly occurs in nature and in our lives at any time and any place. Although each and every material exhibits triboelectrification, its quantification has not been standardized. A triboelectric series has been qualitatively ranked with regards to triboelectric polarization. Here, we introduce a universal standard method to quantify the triboelectric series for a wide range of polymers, establishing quantitative triboelectrification as a fundamental materials property. By measuring the tested materials with a liquid metal in an environment under well-defined conditions, the proposed method standardizes the experimental set up for uniformly quantifying the surface triboelectrification of general materials. The normalized triboelectric charge density is derived to reveal the intrinsic character of polymers for gaining or losing electrons. This quantitative triboelectric series may serve as a textbook standard for implementing the application of triboelectrification for energy harvesting and self-powered sensing.

Habitat: A Platform for Embodied AI Research

Abstract: We present Habitat, a platform for research in embodied artificial intelligence (AI). Habitat enables training embodied agents (virtual robots) in highly efficient photorealistic 3D simulation. Specifically, Habitat consists of: (i) Habitat-Sim: a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling. Habitat-Sim is fast -- when rendering a scene from Matterport3D, it achieves several thousand frames per second (fps) running single-threaded, and can reach over 10,000 fps multi-process on a single GPU. (ii) Habitat-API: a modular high-level library for end-to-end development of embodied AI algorithms -- defining tasks (e.g., navigation, instruction following, question answering), configuring, training, and benchmarking embodied agents. These large-scale engineering contributions enable us to answer scientific questions requiring experiments that were till now impracticable or 'merely' impractical. Specifically, in the context of point-goal navigation: (1) we revisit the comparison between learning and SLAM approaches from two recent works and find evidence for the opposite conclusion -- that learning outperforms SLAM if scaled to an order of magnitude more experience than previous investigations, and (2) we conduct the first cross-dataset generalization experiments {train, test} x {Matterport3D, Gibson} for multiple sensors {blind, RGB, RGBD, D} and find that only agents with depth (D) sensors generalize across datasets. We hope that our open-source platform and these findings will advance research in embodied AI.

First M87 Event Horizon Telescope Results. II. Array and Instrumentation

Abstract: The Event Horizon Telescope (EHT) is a very long baseline interferometry (VLBI) array that comprises millimeter- and submillimeter-wavelength telescopes separated by distances comparable to the diameter of the Earth. At a nominal operating wavelength of ~1.3 mm, EHT angular resolution (λ/D) is ~25 μas, which is sufficient to resolve nearby supermassive black hole candidates on spatial and temporal scales that correspond to their event horizons. With this capability, the EHT scientific goals are to probe general relativistic effects in the strong-field regime and to study accretion and relativistic jet formation near the black hole boundary. In this Letter we describe the system design of the EHT, detail the technology and instrumentation that enable observations, and provide measures of its performance. Meeting the EHT science objectives has required several key developments that have facilitated the robust extension of the VLBI technique to EHT observing wavelengths and the production of instrumentation that can be deployed on a heterogeneous array of existing telescopes and facilities. To meet sensitivity requirements, high-bandwidth digital systems were developed that process data at rates of 64 gigabit s^(−1), exceeding those of currently operating cm-wavelength VLBI arrays by more than an order of magnitude. Associated improvements include the development of phasing systems at array facilities, new receiver installation at several sites, and the deployment of hydrogen maser frequency standards to ensure coherent data capture across the array. These efforts led to the coordination and execution of the first Global EHT observations in 2017 April, and to event-horizon-scale imaging of the supermassive black hole candidate in M87.

Tuning element distribution, structure and properties by composition in high-entropy alloys.

Abstract: High-entropy alloys are a class of materials that contain five or more elements in near-equiatomic proportions1,2. Their unconventional compositions and chemical structures hold promise for achieving unprecedented combinations of mechanical properties3–8. Rational design of such alloys hinges on an understanding of the composition–structure–property relationships in a near-infinite compositional space9,10. Here we use atomic-resolution chemical mapping to reveal the element distribution of the widely studied face-centred cubic CrMnFeCoNi Cantor alloy2 and of a new face-centred cubic alloy, CrFeCoNiPd. In the Cantor alloy, the distribution of the five constituent elements is relatively random and uniform. By contrast, in the CrFeCoNiPd alloy, in which the palladium atoms have a markedly different atomic size and electronegativity from the other elements, the homogeneity decreases considerably; all five elements tend to show greater aggregation, with a wavelength of incipient concentration waves11,12 as small as 1 to 3 nanometres. The resulting nanoscale alternating tensile and compressive strain fields lead to considerable resistance to dislocation glide. In situ transmission electron microscopy during straining experiments reveals massive dislocation cross-slip from the early stage of plastic deformation, resulting in strong dislocation interactions between multiple slip systems. These deformation mechanisms in the CrFeCoNiPd alloy, which differ markedly from those in the Cantor alloy and other face-centred cubic high-entropy alloys, are promoted by pronounced fluctuations in composition and an increase in stacking-fault energy, leading to higher yield strength without compromising strain hardening and tensile ductility. Mapping atomic-scale element distributions opens opportunities for understanding chemical structures and thus providing a basis for tuning composition and atomic configurations to obtain outstanding mechanical properties. In high-entropy alloys, atomic-resolution chemical mapping shows that swapping some of the atoms for larger, more electronegative elements results in atomic-scale modulations that produce higher yield strength, excellent strain hardening and ductility.

Creation and analysis of biochemical constraint-based models using the COBRA Toolbox v.3.0

Abstract: Constraint-based reconstruction and analysis (COBRA) provides a molecular mechanistic framework for integrative analysis of experimental molecular systems biology data and quantitative prediction of physicochemically and biochemically feasible phenotypic states. The COBRA Toolbox is a comprehensive desktop software suite of interoperable COBRA methods. It has found widespread application in biology, biomedicine, and biotechnology because its functions can be flexibly combined to implement tailored COBRA protocols for any biochemical network. This protocol is an update to the COBRA Toolbox v.1.0 and v.2.0. Version 3.0 includes new methods for quality-controlled reconstruction, modeling, topological analysis, strain and experimental design, and network visualization, as well as network integration of chemoinformatic, metabolomic, transcriptomic, proteomic, and thermochemical data. New multi-lingual code integration also enables an expansion in COBRA application scope via high-precision, high-performance, and nonlinear numerical optimization solvers for multi-scale, multi-cellular, and reaction kinetic modeling, respectively. This protocol provides an overview of all these new features and can be adapted to generate and analyze constraint-based models in a wide variety of scenarios. The COBRA Toolbox v.3.0 provides an unparalleled depth of COBRA methods.

Large Intelligent Surface-Assisted Wireless Communication Exploiting Statistical CSI

Abstract: Large intelligent surface (LIS)-assisted wireless communications have drawn attention worldwide. With the use of low-cost LIS on building walls, signals can be reflected by the LIS and sent out along desired directions by controlling its phases, thereby providing supplementary links for wireless communication systems. In this paper, we evaluate the performance of an LIS-assisted large-scale antenna system by formulating a tight upper bound of the ergodic spectral efficiency and investigate the effect of the phase shifts on the ergodic spectral efficiency in different propagation scenarios. In particular, we propose an optimal phase shift design based on the upper bound of the ergodic spectral efficiency and statistical channel state information. Furthermore, we derive the requirement on the quantization bits of the LIS to promise an acceptable spectral efficiency degradation. Numerical results show that using the proposed phase shift design can achieve the maximum ergodic spectral efficiency, and a 2-bit quantizer is sufficient to ensure spectral efficiency degradation of no more than 1 bit/s/Hz.

Properties of the Binary Neutron Star Merger GW170817

Abstract: On August 17, 2017, the Advanced LIGO and Advanced Virgo gravitational-wave detectors observed a low-mass compact binary inspiral. The initial sky localization of the source of the gravitational-wave signal, GW170817, allowed electromagnetic observatories to identify NGC 4993 as the host galaxy. In this work, we improve initial estimates of the binary's properties, including component masses, spins, and tidal parameters, using the known source location, improved modeling, and recalibrated Virgo data. We extend the range of gravitational-wave frequencies considered down to 23 Hz, compared to 30 Hz in the initial analysis. We also compare results inferred using several signal models, which are more accurate and incorporate additional physical effects as compared to the initial analysis. We improve the localization of the gravitational-wave source to a 90% credible region of 16 deg2. We find tighter constraints on the masses, spins, and tidal parameters, and continue to find no evidence for nonzero component spins. The component masses are inferred to lie between 1.00 and 1.89 M when allowing for large component spins, and to lie between 1.16 and 1.60 M (with a total mass 2.73-0.01+0.04 M) when the spins are restricted to be within the range observed in Galactic binary neutron stars. Using a precessing model and allowing for large component spins, we constrain the dimensionless spins of the components to be less than 0.50 for the primary and 0.61 for the secondary. Under minimal assumptions about the nature of the compact objects, our constraints for the tidal deformability parameter Λ are (0,630) when we allow for large component spins, and 300-230+420 (using a 90% highest posterior density interval) when restricting the magnitude of the component spins, ruling out several equation-of-state models at the 90% credible level. Finally, with LIGO and GEO600 data, we use a Bayesian analysis to place upper limits on the amplitude and spectral energy density of a possible postmerger signal.

Recent advances in metal sulfides: from controlled fabrication to electrocatalytic, photocatalytic and photoelectrochemical water splitting and beyond

Abstract: In recent years, nanocrystals of metal sulfide materials have attracted scientific research interest for renewable energy applications due to the abundant choice of materials with easily tunable electronic, optical, physical and chemical properties. Metal sulfides are semiconducting compounds where sulfur is an anion associated with a metal cation; and the metal ions may be in mono-, bi- or multi-form. The diverse range of available metal sulfide materials offers a unique platform to construct a large number of potential materials that demonstrate exotic chemical, physical and electronic phenomena and novel functional properties and applications. To fully exploit the potential of these fascinating materials, scalable methods for the preparation of low-cost metal sulfides, heterostructures, and hybrids of high quality must be developed. This comprehensive review indicates approaches for the controlled fabrication of metal sulfides and subsequently delivers an overview of recent progress in tuning the chemical, physical, optical and nano- and micro-structural properties of metal sulfide nanocrystals using a range of material fabrication methods. For hydrogen energy production, three major approaches are discussed in detail: electrocatalytic hydrogen generation, powder photocatalytic hydrogen generation and photoelectrochemical water splitting. A variety of strategies such as structural tuning, composition control, doping, hybrid structures, heterostructures, defect control, temperature effects and porosity effects on metal sulfide nanocrystals are discussed and how they are exploited to enhance performance and develop future energy materials. From this literature survey, energy conversion currently relies on a limited range of metal sulfides and their composites, and several metal sulfides are immature in terms of their dissolution, photocorrosion and long-term durability in electrolytes during water splitting. Future research directions for innovative metal sulfides should be closely allied to energy and environmental issues, along with their advanced characterization, and developing new classes of metal sulfide materials with well-defined fabrication methods.

Amphibian fungal panzootic causes catastrophic and ongoing loss of biodiversity

Abstract: Anthropogenic trade and development have broken down dispersal barriers, facilitating the spread of diseases that threaten Earth's biodiversity. We present a global, quantitative assessment of the amphibian chytridiomycosis panzootic, one of the most impactful examples of disease spread, and demonstrate its role in the decline of at least 501 amphibian species over the past half-century, including 90 presumed extinctions. The effects of chytridiomycosis have been greatest in large-bodied, range-restricted anurans in wet climates in the Americas and Australia. Declines peaked in the 1980s, and only 12% of declined species show signs of recovery, whereas 39% are experiencing ongoing decline. There is risk of further chytridiomycosis outbreaks in new areas. The chytridiomycosis panzootic represents the greatest recorded loss of biodiversity attributable to a disease.

A Closer Look at Few-shot Classification

Abstract: Few-shot classification aims to learn a classifier to recognize unseen classes during training with limited labeled examples. While significant progress has been made, the growing complexity of network designs, meta-learning algorithms, and differences in implementation details make a fair comparison difficult. In this paper, we present 1) a consistent comparative analysis of several representative few-shot classification algorithms, with results showing that deeper backbones significantly reduce the performance differences among methods on datasets with limited domain differences, 2) a modified baseline method that surprisingly achieves competitive performance when compared with the state-of-the-art on both the \miniI and the CUB datasets, and 3) a new experimental setting for evaluating the cross-domain generalization ability for few-shot classification algorithms. Our results reveal that reducing intra-class variation is an important factor when the feature backbone is shallow, but not as critical when using deeper backbones. In a realistic cross-domain evaluation setting, we show that a baseline method with a standard fine-tuning practice compares favorably against other state-of-the-art few-shot learning algorithms.

Optoelectronic resistive random access memory for neuromorphic vision sensors.

Abstract: Neuromorphic visual systems have considerable potential to emulate basic functions of the human visual system even beyond the visible light region. However, the complex circuitry of artificial visual systems based on conventional image sensors, memory and processing units presents serious challenges in terms of device integration and power consumption. Here we show simple two-terminal optoelectronic resistive random access memory (ORRAM) synaptic devices for an efficient neuromorphic visual system that exhibit non-volatile optical resistive switching and light-tunable synaptic behaviours. The ORRAM arrays enable image sensing and memory functions as well as neuromorphic visual pre-processing with an improved processing efficiency and image recognition rate in the subsequent processing tasks. The proof-of-concept device provides the potential to simplify the circuitry of a neuromorphic visual system and contribute to the development of applications in edge computing and the internet of things.

Control Barrier Functions: Theory and Applications

Abstract: This paper provides an introduction and overview of recent work on control barrier functions and their use to verify and enforce safety properties in the context of (optimization based) safety-critical controllers. We survey the main technical results and discuss applications to several domains including robotic systems.

A Hybrid Approach to Privacy-Preserving Federated Learning

Abstract: Federated learning facilitates the collaborative training of models without the sharing of raw data. However, recent attacks demonstrate that simply maintaining data locality during training processes does not provide sufficient privacy guarantees. Rather, we need a federated learning system capable of preventing inference over both the messages exchanged during training and the final trained model while ensuring the resulting model also has acceptable predictive accuracy. Existing federated learning approaches either use secure multiparty computation (SMC) which is vulnerable to inference or differential privacy which can lead to low accuracy given a large number of parties with relatively small amounts of data each. In this paper, we present an alternative approach that utilizes both differential privacy and SMC to balance these trade-offs. Combining differential privacy with secure multiparty computation enables us to reduce the growth of noise injection as the number of parties increases without sacrificing privacy while maintaining a pre-defined rate of trust. Our system is therefore a scalable approach that protects against inference threats and produces models with high accuracy. Additionally, our system can be used to train a variety of machine learning models, which we validate with experimental results on 3 different machine learning algorithms. Our experiments demonstrate that our approach out-performs state of the art solutions.

Attention is not not Explanation

Abstract: Attention mechanisms play a central role in NLP systems, especially within recurrent neural network (RNN) models. Recently, there has been increasing interest in whether or not the intermediate representations offered by these modules may be used to explain the reasoning for a model’s prediction, and consequently reach insights regarding the model’s decision-making process. A recent paper claims that ‘Attention is not Explanation’ (Jain and Wallace, 2019). We challenge many of the assumptions underlying this work, arguing that such a claim depends on one’s definition of explanation, and that testing it needs to take into account all elements of the model. We propose four alternative tests to determine when/whether attention can be used as explanation: a simple uniform-weights baseline; a variance calibration based on multiple random seed runs; a diagnostic framework using frozen weights from pretrained models; and an end-to-end adversarial attention training protocol. Each allows for meaningful interpretation of attention mechanisms in RNN models. We show that even when reliable adversarial distributions can be found, they don’t perform well on the simple diagnostic, indicating that prior work does not disprove the usefulness of attention mechanisms for explainability.

Visual Dialog

Abstract: We introduce the task of Visual Dialog, which requires an AI agent to hold a meaningful dialog with humans in natural, conversational language about visual content. Specifically, given an image, a dialog history, and a question about the image, the agent has to ground the question in image, infer context from history, and answer the question accurately. Visual Dialog is disentangled enough from a specific downstream task so as to serve as a general test of machine intelligence, while being sufficiently grounded in vision to allow objective evaluation of individual responses and benchmark progress. We develop a novel two-person real-time chat data-collection protocol to curate a large-scale Visual Dialog dataset (VisDial). VisDial v0.9 has been released and consists of $\sim$ ∼ 1.2M dialog question-answer pairs from 10-round, human-human dialogs grounded in $\sim$ ∼ 120k images from the COCO dataset. We introduce a family of neural encoder-decoder models for Visual Dialog with 3 encoders—Late Fusion, Hierarchical Recurrent Encoder and Memory Network (optionally with attention over image features)—and 2 decoders (generative and discriminative), which outperform a number of sophisticated baselines. We propose a retrieval-based evaluation protocol for Visual Dialog where the AI agent is asked to sort a set of candidate answers and evaluated on metrics such as mean-reciprocal-rank and recall $@k$ @ k of human response. We quantify the gap between machine and human performance on the Visual Dialog task via human studies. Putting it all together, we demonstrate the first ‘visual chatbot’! Our dataset, code, pretrained models and visual chatbot are available on https://visualdialog.org .

Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers

Abstract: Deep learning has recently seen rapid development and received significant attention due to its state-of-the-art performance on previously-thought hard problems. However, because of the internal complexity and nonlinear structure of deep neural networks, the underlying decision making processes for why these models are achieving such performance are challenging and sometimes mystifying to interpret. As deep learning spreads across domains, it is of paramount importance that we equip users of deep learning with tools for understanding when a model works correctly, when it fails, and ultimately how to improve its performance. Standardized toolkits for building neural networks have helped democratize deep learning; visual analytics systems have now been developed to support model explanation, interpretation, debugging, and improvement. We present a survey of the role of visual analytics in deep learning research, which highlights its short yet impactful history and thoroughly summarizes the state-of-the-art using a human-centered interrogative framework, focusing on the Five W's and How (Why, Who, What, How, When, and Where). We conclude by highlighting research directions and open research problems. This survey helps researchers and practitioners in both visual analytics and deep learning to quickly learn key aspects of this young and rapidly growing body of research, whose impact spans a diverse range of domains.

A Closer Look at Few-shot Classification

Abstract: Few-shot classification aims to learn a classifier to recognize unseen classes during training with limited labeled examples. While significant progress has been made, the growing complexity of network designs, meta-learning algorithms, and differences in implementation details make a fair comparison difficult. In this paper, we present 1) a consistent comparative analysis of several representative few-shot classification algorithms, with results showing that deeper backbones significantly reduce the gap across methods when domain differences are limited, 2) a slightly modified baseline method that surprisingly achieves competitive performance when compared with the state-of-the-art on both the mini-ImageNet and the CUB datasets, and 3) a new experimental setting for evaluating the cross-domain generalization ability for few-shot classification algorithms. Our results reveal that reducing intra-class variation is an important factor when the feature backbone is shallow, but not as critical when using deeper backbones. In a realistic, cross-domain evaluation setting, we show that a baseline method with a standard fine-tuning practice compares favorably against other state-of-the-art few-shot learning algorithms.

Sodium Metal Anodes: Emerging Solutions to Dendrite Growth

Abstract: This comprehensive Review focuses on the key challenges and recent progress regarding sodium-metal anodes employed in sodium-metal batteries (SMBs) The metal anode is the essential component of emerging energy storage systems such as sodium sulfur and sodium selenium, which are discussed as example full-cell applications We begin with a description of the differences in the chemical and physical properties of Na metal versus the oft-studied Li metal, and a corresponding discussion regarding the number of ways in which Na does not follow Li-inherited paradigms in its electrochemical behavior We detail the major challenges for Na-metal systems that at this time limit the feasibility of SMBs The core Na anode problems are the following interrelated degradation mechanisms: An unstable solid electrolyte interphase with most organic electrolytes, "mossy" and "lath-like" metal dendrite growth for liquid systems, poor Coulombic efficiency, and gas evolution Even solid-state Na batteries are not immune, with metal dendrites being reported The solutions may be subdivided into the following interrelated taxonomy: Improved electrolytes and electrolyte additives tailored for Na-metal anodes, interfacial engineering between the metal and the liquid or solid electrolyte, electrode architectures that both reduce the current density during plating-stripping and serve as effective hosts that shield the Na metal from excessive reactions, and alloy design to tune the bulk properties of the metal per se For instance, stable plating-stripping of Na is extremely difficult with conventional carbonate solvents but has been reported with ethers and glymes Solid-state electrolytes (SSEs) such as beta-alumina solid electrolyte (BASE), sodium superionic conductor (NASICON), and sodium thiophosphate (75Na2S·25P2S5) present highly exciting opportunities for SMBs that avoid the dangers of flammable liquids Even SSEs are not immune to dendrites, however, which grow through the defects in the bulk pellet, but may be controlled through interfacial energy modification We conclude with a discussion of the key research areas that we feel are the most fruitful for further pursuit In our opinion, greatly improved understanding and control of the SEI structure is the key to cycling stability A holistic approach involving complementary post-mortem, in situ, and operando analyses to elucidate full battery cell level structure-performance relations is advocated

Binary Black Hole Population Properties Inferred from the First and Second Observing Runs of Advanced LIGO and Advanced Virgo

Abstract: We present results on the mass, spin, and redshift distributions with phenomenological population models using the 10 binary black hole (BBH) mergers detected in the first and second observing runs completed by Advanced LIGO and Advanced Virgo. We constrain properties of the BBH mass spectrum using models with a range of parameterizations of the BBH mass and spin distributions. We find that the mass distribution of the more massive BH in such binaries is well approximated by models with no more than 1% of BHs more massive than 45 M and a power-law index of (90% credibility). We also show that BBHs are unlikely to be composed of BHs with large spins aligned to the orbital angular momentum. Modeling the evolution of the BBH merger rate with redshift, we show that it is flat or increasing with redshift with 93% probability. Marginalizing over uncertainties in the BBH population, we find robust estimates of the BBH merger rate density of R= (90% credibility). As the BBH catalog grows in future observing runs, we expect that uncertainties in the population model parameters will shrink, potentially providing insights into the formation of BHs via supernovae, binary interactions of massive stars, stellar cluster dynamics, and the formation history of BHs across cosmic time.