arXiv subjects

Yi Ren

Publications and source records attributed to Yi Ren.

At least 253 records · Page 14Linked to original sources

arXiv2019

Crystallographic Characterization of Black Phosphorene and Its Application in Nanostructures

The central question in the field of 2D materials is how a material behaves when it is patterned at nanometer scale with different edge geometries. Due to the anisotropy inherent in the puckered structure, black phosphorene nanostructures may have more varieties of edge geometries. Here, we present a comprehensive 2D planar crystallographic characterization of phosphorene uniformly by a chiral vector (angle), from which a new type of edge atomic configuration, the slope edge geometry is discovered. The phosphorene nanoribbons (PNRs) with slope edges, like previously noticed zigzag and skewed-armchair PNRs, also own interesting twofold-degenerate edge states. These three marginal directions, together with the skewed-zigzag and armchair directions without edge states, divide a phosphorene into four regions among which the electronic property is different from each other. Further, for a PNR cutting along any possible direction, by taking into account the z-direction in structure, whether or not it owns edge states depends on the periodic zigzag-like morphology along the edges. Moreover, we propose a PNR-based z-shaped homogenous junction with scale $\sim$100 nm, where the central scattering region between two PNR electrodes is a phosphorene quantum dot (PQD) of PNR segment with various edge morphologies. Interestingly, the calculated conductance of the junction relies sensitively on the central PQD edge states, which exhibits a terrace with resonant oscillations near Fermi energy, otherwise the electron transport is blocked due to the absence of edge states. Remarkably, the number of oscillating peaks exactly matches to the number of sawtooth along an edge of PQD, since the discrete energy levels of the zigzag-like edge provide the transporting channels for electrons. The results here may be extended to the group-VA 2D materials for observing the electronic property of nanostructures.

cond-mat.mes-hall↗

arXiv2019

Form + Function: Optimizing Aesthetic Product Design via Adaptive, Geometrized Preference Elicitation

Visual design is critical to product success, and the subject of intensive marketing research effort. Yet visual elements, due to their holistic and interactive nature, do not lend themselves well to optimization using extant decompositional methods for preference elicitation. Here we present a systematic methodology to incorporate interactive, 3D-rendered product configurations into a conjoint-like framework. The method relies on rapid, scalable machine learning algorithms to adaptively update product designs along with standard information-oriented product attributes. At its heart is a parametric account of a product's geometry, along with a novel, adaptive "bi-level" query task that can estimate individuals' visual design form preferences and their trade-offs against such traditional elements as price and product features. We illustrate the method's performance through extensive simulations and robustness checks, a formal proof of the bi-level query methodology's domain of superiority, and a field test for the design of a mid-priced sedan, using real-time 3D rendering for an online panel. Results indicate not only substantially enhanced predictive accuracy, but two quantities beyond the reach of standard conjoint methods: trade-offs between form and function overall, and willingness-to-pay for specific design elements. Moreover -- and most critically for applications -- the method provides "optimal" visual designs for both individuals and model-derived or analyst-supplied consumer groupings, as well as their sensitivities to form and functional elements.

cs.HC↗

arXiv2019

FastSpeech: Fast, Robust and Controllable Text to Speech

Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.

cs.CL↗

arXiv2019

The Emergence of Compositional Languages for Numeric Concepts Through Iterated Learning in Neural Agents

Since first introduced, computer simulation has been an increasingly important tool in evolutionary linguistics. Recently, with the development of deep learning techniques, research in grounded language learning has also started to focus on facilitating the emergence of compositional languages without pre-defined elementary linguistic knowledge. In this work, we explore the emergence of compositional languages for numeric concepts in multi-agent communication systems. We demonstrate that compositional language for encoding numeric concepts can emerge through iterated learning in populations of deep neural network agents. However, language properties greatly depend on the input representations given to agents. We found that compositional languages only emerge if they require less iterations to be fully learnt than other non-degenerate languages for agents on a given input representation.

cs.CL↗

arXiv2019

Evolved Massive Stars at Low-metallicity I. A Source Catalog for the Small Magellanic Cloud

We present a clean, magnitude-limited (IRAC1 or WISE1 $\leq$ 15.0 mag) multiwavelength source catalog for the SMC with 45,466 targets in total, with the purpose of building an anchor for future studies, especially for the massive star populations at low-metallicity. The catalog contains data in 50 different bands including 21 optical and 29 infrared bands, ranging from the ultraviolet to the far-infrared. Additionally, radial velocities and spectral classifications were collected from the literature, as well as infrared and optical variability statistics were retrieved from different projects. The catalog was essentially built upon a $1''$ crossmatching and a $3''$ deblending between the SEIP source list and Gaia DR2 photometric data. Further constraints on the proper motions and parallaxes from Gaia DR2 allowed us to remove the foreground contamination. We estimated that about 99.5\% of the targets in our catalog were most likely genuine members of the SMC. By using the evolutionary tracks and synthetic photometry from MIST and the theoretical $\rm J-K_S$ color cuts, we identified 1,405 RSG, 217 YSG and 1,369 BSG candidates in the SMC in five different CMDs, where attention should also be paid to the incompleteness of our sample. We ranked the candidates based on the intersection of different CMDs. A comparison between the models and observational data shows that the lower limit of initial mass for the RSGs population may be as low as 7 or even 6 $M_{\odot}$ and the RSG is well separated from the AGB population even at faint magnitude, making RSGs a unique population connecting the evolved massive and intermediate stars, since stars with initial mass around 6 to 8 $M_{\odot}$ are thought to go through a second dredge-up to become AGBs. We encourage the interested reader to further exploit the potential of our catalog.

astro-ph.SR↗

arXiv2019

Augmenting Model Robustness with Transformation-Invariant Attacks

The vulnerability of neural networks under adversarial attacks has raised serious concerns and motivated extensive research. It has been shown that both neural networks and adversarial attacks against them can be sensitive to input transformations such as linear translation and rotation, and that human vision, which is robust against adversarial attacks, is invariant to natural input transformations. Based on these, this paper tests the hypothesis that model robustness can be further improved when it is adversarially trained against transformed attacks and transformation-invariant attacks. Experiments on MNIST, CIFAR-10, and restricted ImageNet show that while transformations of attacks alone do not affect robustness, transformation-invariant attacks can improve model robustness by 2.5\% on MNIST, 3.7\% on CIFAR-10, and 1.1\% on restricted ImageNet. We discuss the intuition behind this phenomenon.

cs.CV↗

arXiv2019

Efficient Subsampled Gauss-Newton and Natural Gradient Methods for Training Neural Networks

We present practical Levenberg-Marquardt variants of Gauss-Newton and natural gradient methods for solving non-convex optimization problems that arise in training deep neural networks involving enormous numbers of variables and huge data sets. Our methods use subsampled Gauss-Newton or Fisher information matrices and either subsampled gradient estimates (fully stochastic) or full gradients (semi-stochastic), which, in the latter case, we prove convergent to a stationary point. By using the Sherman-Morrison-Woodbury formula with automatic differentiation (backpropagation) we show how our methods can be implemented to perform efficiently. Finally, numerical results are presented to demonstrate the effectiveness of our proposed methods.

cs.LG↗

arXiv2019

Novel Hierarchical Correlation Functions for Quantitative Representation of Complex Heterogeneous Materials and Microstructural Evolution

Effective and accurate characterization and quantification of complex microstructure of a heterogeneous material and its evolution under external stimuli are very challenging, yet crucial to achieving reliable material performance prediction, processing optimization and advanced material design. Here, we address this challenge by developing a set of novel hierarchical statistical microstructural descriptors, which we call the "n-point polytope functions" Pn, for quantitative characterization, representation and modeling of complex material microstructure and its evolution. These novel polytope functions successively include higher-order n-point statistics of the features of interest in the microstructure in a concise, expressive, explainable, and universal manner; and can be directly computed from multi-modal imaging data. We develop highly efficient computational tools to directly extract the Pn functions up to n = 8 from multi-modal imaging data. Using simple model microstructures, we show that these novel statistical descriptors effectively "decompose" the structural features of interest into a set of "polytope basis", allowing one to easily detect any underlying symmetry or emerging features during the structural evolution. We apply these novel Pn functions to quantify and model a variety of heterogeneous material systems, including particle-reinforced composites, metal-ceramic composites, concretes, porous materials; as well as the microstructural evolution in an aged lead-tin alloy. Our results indicate that the Pn functions can offer a practically complete and compact set of basis for quantitative microstructure representation (QMR), for both static 3D complex microstructure and 4D microstructural evolution of a wide spectrum of heterogeneous material systems.

cond-mat.mtrl-sci↗

arXiv2019

Multilingual Neural Machine Translation with Knowledge Distillation

Multilingual machine translation, which translates multiple languages with a single model, has attracted much attention due to its efficiency of offline training and online serving. However, traditional multilingual translation usually yields inferior accuracy compared with the counterpart using individual models for each language pair, due to language diversity and model capacity limitations. In this paper, we propose a distillation-based approach to boost the accuracy of multilingual machine translation. Specifically, individual models are first trained and regarded as teachers, and then the multilingual model is trained to fit the training data and match the outputs of individual models simultaneously through knowledge distillation. Experiments on IWSLT, WMT and Ted talk translation datasets demonstrate the effectiveness of our method. Particularly, we show that one model is enough to handle multiple languages (up to 44 languages in our experiment), with comparable or even better accuracy than individual models.

cs.CL↗

arXiv2019

The Period-Luminosity Relations of Red Supergiants in M33 and M31

Based on previously selected preliminary samples of Red Supergiants (RSGs) in M33 and M31, the foreground stars and luminous Asymptotic Giant Branch stars (AGBs) are further excluded, which leads to the samples of 717 RSGs in M33 and 420 RSGs in M31. With the time-series data from the iPTF survey spanning nearly 2000 days, the period and amplitude of RSGs are analyzed. According to the lightcurves characteristics, they are classified into four categories in which 84 and 56 objects in M33 and M31 respectively are semi-regular variables. For these semi-regular variables, the pulsation mode is identified by comparing with the theoretical model, which yielded 19 (7) sources in the first overtone mode in M33 (M31), and the other 65 (49) RSGs in M33 (M31) in the fundamental mode. The period-luminosity (P-L) relation is analyzed for the RSGs in the fundamental mode. It is found the P-L relation is tight in the infrared, i.e. the 2MASS $JHK_{\rm S}$ bands and the short-wavelength bands of Spitzer. Meanwhile, the inhomogeneous extinction causes the P-L relation scattering in the $V$ band, and the dust emission causes the less tight P-L relation in the Spitzer/[8.0] and [24] bands. The derived P-L relations in the 2MASS/$K_{\rm S}$ band are in agreement with those of RSGs in SMC, LMC and the Milky Way within the uncertainty range. It is found that the number ratio of RSGs pulsating in the fundamental mode to the first overtone mode increases with metallicity.

astro-ph.SR↗

arXiv2019

Image Decomposition and Classification through a Generative Model

We demonstrate in this paper that a generative model can be designed to perform classification tasks under challenging settings, including adversarial attacks and input distribution shifts. Specifically, we propose a conditional variational autoencoder that learns both the decomposition of inputs and the distributions of the resulting components. During test, we jointly optimize the latent variables of the generator and the relaxed component labels to find the best match between the given input and the output of the generator. The model demonstrates promising performance at recognizing overlapping components from the multiMNIST dataset, and novel component combinations from a traffic sign dataset. Experiments also show that the proposed model achieves high robustness on MNIST and NORB datasets, in particular for high-strength gradient attacks and non-gradient attacks.

cs.CV↗

arXiv2019

Low-cost Measurement of Industrial Shock Signals via Deep Learning Calibration

Special high-end sensors with expensive hardware are usually needed to measure shock signals with high accuracy. In this paper, we show that cheap low-end sensors calibrated by deep neural networks are also capable to measure high-g shocks accurately. Firstly we perform drop shock tests to collect a dataset of shock signals measured by sensors of different fidelity. Secondly, we propose a novel network to effectively learn both the signal peak and overall shape. The results show that the proposed network is capable to map low-end shock signals to its high-end counterparts with satisfactory accuracy. To the best of our knowledge, this is the first work to apply deep learning techniques to calibrate shock sensors.

eess.SP↗

arXiv2019

Conventional and unconventional photon blockade effects in an atom-cavity system

A two-level system interacting with a cavity field is an important model for investigating the photon blockade (PB) effect. Most work on this topic has been based on the assumption that the atomic transition frequency is resonant with the fundamental mode frequency of the cavity. We relax this constraint and reexamine PB in a more general atom--cavity system with arbitrary atomic and cavity detunings from a driving field. The results show that when the signs of the atomic and cavity detunings are the same, PB occurs only in the strong-coupling regime, but for opposite signs of the atomic and cavity detunings, strong photon antibunching is observed in both the weak- and strong-coupling regimes and a better PB effect is achieved compared with the case when the signs are the same. More interestingly, we find that this PB arises from quantum interference for both weak and strong nonlinearities. These results deepen our understanding of the underlying mechanism of PB and may be help in the construction of single-photon sources with higher purity and better flexibility using atom--cavity systems.

quant-ph↗

arXiv2019

How Shall I Drive? Interaction Modeling and Motion Planning towards Empathetic and Socially-Graceful Driving

While intelligence of autonomous vehicles (AVs) has significantly advanced in recent years, accidents involving AVs suggest that these autonomous systems lack gracefulness in driving when interacting with human drivers. In the setting of a two-player game, we propose model predictive control based on social gracefulness, which is measured by the discrepancy between the actions taken by the AV and those that could have been taken in favor of the human driver. We define social awareness as the ability of an agent to infer such favorable actions based on knowledge about the other agent's intent, and further show that empathy, i.e., the ability to understand others' intent by simultaneously inferring others' understanding of the agent's self intent, is critical to successful intent inference. Lastly, through an intersection case, we show that the proposed gracefulness objective allows an AV to learn more sophisticated behavior, such as passive-aggressive motions that gently force the other agent to yield.

cs.RO↗

arXiv2018

One-Shot Generation of Near-Optimal Topology through Theory-Driven Machine Learning

We introduce a theory-driven mechanism for learning a neural network model that performs generative topology design in one shot given a problem setting, circumventing the conventional iterative process that computational design tasks usually entail. The proposed mechanism can lead to machines that quickly response to new design requirements based on its knowledge accumulated through past experiences of design generation. Achieving such a mechanism through supervised learning would require an impractically large amount of problem-solution pairs for training, due to the known limitation of deep neural networks in knowledge generalization. To this end, we introduce an interaction between a student (the neural network) and a teacher (the optimality conditions underlying topology optimization): The student learns from existing data and is tested on unseen problems. Deviation of the student's solutions from the optimality conditions is quantified, and used for choosing new data points to learn from. We call this learning mechanism "theory-driven", as it explicitly uses domain-specific theories to guide the learning, thus distinguishing itself from purely data-driven supervised learning. We show through a compliance minimization problem that the proposed learning mechanism leads to topology generation with near-optimal structural compliance, much improved from standard supervised learning under the same computational budget.

cs.LG↗

arXiv2018

Gear Training: A new way to implement high-performance model-parallel training

The training of Deep Neural Networks usually needs tremendous computing resources. Therefore many deep models are trained in large cluster instead of single machine or GPU. Though major researchs at present try to run whole model on all machines by using asynchronous asynchronous stochastic gradient descent (ASGD), we present a new approach to train deep model parallely -- split the model and then seperately train different parts of it in different speed.

cs.LG↗

arXiv2018

Liquid Metal Enabled Electrobiology: A Generalized Easy Going Way to Tackle Disease Challenges

In this article, a new conceptual biomedical engineering strategy to tackle modern disease challenges, termed as liquid metal enabled electrobiology, is proposed. This generalized and easy going way is based on the physiological fact that specially administrated electricity would induce a series of subsequent desired biological effects, either shortly, transitional or permanently. Owing high compliance within any part of the biological tissues, the liquid metal would aid to mold a pervasive way to treat physiological or psychological diseases. As highly conductive and non-toxic multifunctional flexible materials, such liquid metals (LMs) consist of the core to generate any requested electric treating fields (ETFields) which can adapt to various sites inside the human body. The basic mechanisms of electrobiology in delivering electricity to the target tissues and then inducing expected outputs for disease treatment are interpreted. The methods toward realizing soft and conformable electronics based on liquid metal are illustrated. Further, a group of typical disease challenges were taken to illustrate the basic strategies to perform liquid metal electrobiology therapy, which include but are not limited to: tissue electronics, brain disorder, immunotherapy, neural functional recovery, muscle stimulation, skin rejuvenation, cosmetology and dieting, artificial organs, cardiac pacing and cancer therapy etc. Some practical issues involved in the electrobiology for future disease therapy were discussed. Perspective along this direction to incubate an easy-going biomedical tool for health care was pointed out.

physics.med-ph↗

arXiv2018

Red supergiant stars in the Large Magellanic Cloud. II. Infrared properties and mid-infrared variability

The characteristics of IR properties and MIR variability of RSGs in the LMC are analyzed based on 12 bands of NIR to MIR co-added data from 2MASS, Spitzer and WISE, and $\sim$6.6 years of MIR time-series data collected by the ALLWISE and NEOWISE-R projects. 773 RSGs candidates were compiled from the literature and verified by using the CMD, SED and MIR variability. About 15\% of valid targets in the $IRAC1-IRAC2$/$IRAC2-IRAC3$ diagram may show PAH emission. We show that arbitrary dereddening Q parameters related to the IRAC4, S9W, WISE3, WISE4, and MIPS24 bands could be constructed based on a precise measurement of MIR interstellar extinction law. Several peculiar outliers in our sample are discussed, in which one outlier might be a RSG right before the explosion or an x-AGB star in the very late evolutionary stage based on the MIR spectrum and photometry. There are 744 identified RSGs in the final sample having both the WISE1- and WISE2-band time-series data. The results show that the MIR variability is increasing along with the increasing of brightness. There is a relatively tight correlation between the MIR variability, MLR, and the warm dust or continuum, where the MIR variability is evident for the targets with $K_S-WISE3>1.0~mag$ and $WISE4<6.5~mag$, while the rest of the targets show much smaller MIR variability. The MIR variability is also correlated with the MLR for which targets with larger variability also show larger MLR with an approximate upper limit of $-6.1~M_\odot/yr^{-1}$. Both the variability and the luminosity may be important for the MLR since the WISE4-band flux is increasing exponentially along with the degeneracy of luminosity and variability. The identified RSG sample has been compared with the theoretical evolutionary models and shown that the discrepancy between observation and evolutionary models can be mitigated by considering both variability and extinction.

astro-ph.SR↗