arXiv subjects

Ziwei Wang

Publications and source records attributed to Ziwei Wang.

At least 217 records · Page 12Linked to original sources

arXiv2020

ORD: Object Relationship Discovery for Visual Dialogue Generation

With the rapid advancement of image captioning and visual question answering at single-round level, the question of how to generate multi-round dialogue about visual content has not yet been well explored.Existing visual dialogue methods encode the image into a fixed feature vector directly, concatenated with the question and history embeddings to predict the response.Some recent methods tackle the co-reference resolution problem using co-attention mechanism to cross-refer relevant elements from the image, history, and the target question.However, it remains challenging to reason visual relationships, since the fine-grained object-level information is omitted before co-attentive reasoning. In this paper, we propose an object relationship discovery (ORD) framework to preserve the object interactions for visual dialogue generation. Specifically, a hierarchical graph convolutional network (HierGCN) is proposed to retain the object nodes and neighbour relationships locally, and then refines the object-object connections globally to obtain the final graph embeddings. A graph attention is further incorporated to dynamically attend to this graph-structured representation at the response reasoning stage. Extensive experiments have proved that the proposed method can significantly improve the quality of dialogue by utilising the contextual information of visual relationships. The model achieves superior performance over the state-of-the-art methods on the Visual Dialog dataset, increasing MRR from 0.6222 to 0.6447, and recall@1 from 48.48% to 51.22%.

cs.CV↗

arXiv2020

Light Dark Photon Dark Matter from Inflation

We discuss the possibility of producing a light dark photon dark matter through a coupling between the dark photon field and the inflaton. The dark photon with a large wavelength is efficiently produced due to the inflaton motion during inflation and becomes non-relativistic before the time of matter-radiation equality. We compute the amount of production analytically. The correct relic abundance is realized with a dark photon mass extending down to $10^{-21} \, \rm eV$.

hep-ph↗

arXiv2020

Nonsingular Ekpyrotic Cosmology with a Nearly Scale-Invariant Spectrum of Cosmological Perturbations and Gravitational Waves

We propose a mechanism borrowed from string theory which yields a non-singular transition from a phase of Ekpyrotic contraction to the expanding phase of Standard Big Bang cosmology. The same mechanism converts the initial vacuum spectrum of cosmological fluctuations before the bounce into a scale-invariant one, and also changes the spectrum of gravitational waves into an almost scale-invariant one. The scalar and tensor tilts are predicted to be the same, in contrast to the predictions from the "String Gas Cosmology" scenario. The amplitude of the gravitational wave spectrum depends on the ratio of the string scale to the Planck scale and may be in reach of upcoming experiments.

hep-th↗

arXiv2020

Ekpyrotic Cosmology with a Zero-Shear S-Brane

In a recent paper we proposed a mechanism for a continuous transition between a contracting Ekpyrotic phase and the Standard Big Bang phase of expansion: the bounce is generated by an S-brane which represents the effects of higher mass string states in the low energy effective field theory. We showed that gravitational waves on cosmological scales obtain a nearly scale-invariant spectrum. Here, we study the cosmological fluctuations in this setup, assuming that the S-brane has zero shear. We find a nearly scale-invariant spectrum of cosmological perturbations with a slight red tilt. The scenario yields two consistency relations for cosmological observations, the first one relating the tensor to scalar ratio with the scalar spectral tilt, the second relating the tensor tilt to the scalar tilt. The predicted tensor to scalar ratio is within the reach of upcoming CMB observations. The tensor tilt is blue.

hep-th↗

arXiv2020

The prescribed time sliding mode control for attitude tracking of spacecraft

With the development of the space missions, there are extensive missions in the demand of the prescribed time convergence. However, it is still a difficult work to combine the prescribed time method with the sliding mode control due to the infinite gain of the prescribed time method while approaching the prescribed time and two periods of sliding mode control. In this paper, a new prescribed time sliding mode control method is proposed for general systems with matched disturbances, from the second-order system to the high-order system. A novel sliding mode variable with explicit time term is designed for achieving the prescribed time convergence. More importantly, as time approaches the prescribed time, the singularity of control input can be avoided. Finally, this paper presents a disturbance observer based prescribed time sliding mode control method for attitude tracking of spacecraft and the efficiency of this method has been verified through the numerical simulations.

eess.SY↗

arXiv2020

BiDet: An Efficient Binarized Object Detector

In this paper, we propose a binarized neural network learning method called BiDet for efficient object detection. Conventional network binarization methods directly quantize the weights and activations in one-stage or two-stage detectors with constrained representational capacity, so that the information redundancy in the networks causes numerous false positives and degrades the performance significantly. On the contrary, our BiDet fully utilizes the representational capacity of the binary neural networks for object detection by redundancy removal, through which the detection precision is enhanced with alleviated false positives. Specifically, we generalize the information bottleneck (IB) principle to object detection, where the amount of information in the high-level feature maps is constrained and the mutual information between the feature maps and object detection is maximized. Meanwhile, we learn sparse object priors so that the posteriors are concentrated on informative detection prediction with false positive elimination. Extensive experiments on the PASCAL VOC and COCO datasets show that our method outperforms the state-of-the-art binary neural networks by a sizable margin.

cs.CV↗

arXiv2019

Learning from the Past: Continual Meta-Learning via Bayesian Graph Modeling

Meta-learning for few-shot learning allows a machine to leverage previously acquired knowledge as a prior, thus improving the performance on novel tasks with only small amounts of data. However, most mainstream models suffer from catastrophic forgetting and insufficient robustness issues, thereby failing to fully retain or exploit long-term knowledge while being prone to cause severe error accumulation. In this paper, we propose a novel Continual Meta-Learning approach with Bayesian Graph Neural Networks (CML-BGNN) that mathematically formulates meta-learning as continual learning of a sequence of tasks. With each task forming as a graph, the intra- and inter-task correlations can be well preserved via message-passing and history transition. To remedy topological uncertainty from graph initialization, we utilize Bayes by Backprop strategy that approximates the posterior distribution of task-specific parameters with amortized inference networks, which are seamlessly integrated into the end-to-end edge learning. Extensive experiments conducted on the miniImageNet and tieredImageNet datasets demonstrate the effectiveness and efficiency of the proposed method, improving the performance by 42.8% compared with state-of-the-art on the miniImageNet 5-way 1-shot classification task.

cs.LG↗

arXiv2019

Tunable High-Quality Fano Resonance in Coupled Terahertz Whispering-Gallery-Mode Resonators

Fano resonance is widely discussed in designing novel terahertz components, such as sensors, filters, modulators, and group delay modules. Usually, high quality (Q) factor and flexible tunability of Fano resonance are key requirements for these applications. Here, we present tunable terahertz Fano resonance with a Q factor of 2095 at 0.439 THz in coupled terahertz whispering-gallery-mode resonators (WGMRs). Coupling between a relatively low Q (578) quartz ring and a high Q (2095) silicon ring is employed to generate the Fano resonance. The resonant frequency of the Fano resonance can be actively manipulated by tuning the resonant frequency of the high Q WGMR, which is achieved through utilizing an electrical thermo-optic tuning method, meanwhile, the resonance intensity of the Fano resonance can be engineered by adjusting the coupling strength between two WGMRs. This coupled-WGMR scheme delivers high Q tunable Fano resonance and may contribute to the design of high-performance configurable terahertz devices.

physics.optics↗

arXiv2019

Curiosity-driven Reinforcement Learning for Diverse Visual Paragraph Generation

Visual paragraph generation aims to automatically describe a given image from different perspectives and organize sentences in a coherent way. In this paper, we address three critical challenges for this task in a reinforcement learning setting: the mode collapse, the delayed feedback, and the time-consuming warm-up for policy networks. Generally, we propose a novel Curiosity-driven Reinforcement Learning (CRL) framework to jointly enhance the diversity and accuracy of the generated paragraphs. First, by modeling the paragraph captioning as a long-term decision-making process and measuring the prediction uncertainty of state transitions as intrinsic rewards, the model is incentivized to memorize precise but rarely spotted descriptions to context, rather than being biased towards frequent fragments and generic patterns. Second, since the extrinsic reward from evaluation is only available until the complete paragraph is generated, we estimate its expected value at each time step with temporal-difference learning, by considering the correlations between successive actions. Then the estimated extrinsic rewards are complemented by dense intrinsic rewards produced from the derived curiosity module, in order to encourage the policy to fully explore action space and find a global optimum. Third, discounted imitation learning is integrated for learning from human demonstrations, without separately performing the time-consuming warm-up in advance. Extensive experiments conducted on the Standford image-paragraph dataset demonstrate the effectiveness and efficiency of the proposed method, improving the performance by 38.4% compared with state-of-the-art.

cs.CV↗

arXiv2019

Eternal Inflation, Entropy Bounds and the Swampland

It has been suggested that low energy effective field theories should satisfy given conditions in order to be successfully embedded into string theory. In the case of a single canonically normalized scalar field this translates into conditions on its potential and the derivatives thereof. In this Letter we revisit stochastic models of small field inflation and study the compatibility of the swampland constraints with entropy considerations. We show that stochastic inflation either violates entropy bounds or the swampland criterium on the slope of the scalar field potential. Furthermore, we illustrate that such models are faced with a graceful exit problem: any patch of space which exits the region of eternal inflation is either not large enough to explain the isotropy of the cosmic microwave background, or has a spectrum of fluctuations with an unacceptably large red tilt.

hep-th↗

arXiv2019

Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration

Previous works utilized ''smaller-norm-less-important'' criterion to prune filters with smaller norm values in a convolutional neural network. In this paper, we analyze this norm-based criterion and point out that its effectiveness depends on two requirements that are not always met: (1) the norm deviation of the filters should be large; (2) the minimum norm of the filters should be small. To solve this problem, we propose a novel filter pruning method, namely Filter Pruning via Geometric Median (FPGM), to compress the model regardless of those two requirements. Unlike previous methods, FPGM compresses CNN models by pruning filters with redundancy, rather than those with ''relatively less'' importance. When applied to two image classification benchmarks, our method validates its usefulness and strengths. Notably, on CIFAR-10, FPGM reduces more than 52% FLOPs on ResNet-110 with even 2.69% relative accuracy improvement. Moreover, on ILSVRC-2012, FPGM reduces more than 42% FLOPs on ResNet-101 without top-5 accuracy drop, which has advanced the state-of-the-art. Code is publicly available on GitHub: https://github.com/he-y/filter-pruning-geometric-median

cs.CV↗

arXiv2019

Snap and Find: Deep Discrete Cross-domain Garment Image Retrieval

With the increasing number of online stores, there is a pressing need for intelligent search systems to understand the item photos snapped by customers and search against large-scale product databases to find their desired items. However, it is challenging for conventional retrieval systems to match up the item photos captured by customers and the ones officially released by stores, especially for garment images. To bridge the customer- and store- provided garment photos, existing studies have been widely exploiting the clothing attributes (\textit{e.g.,} black) and landmarks (\textit{e.g.,} collar) to learn a common embedding space for garment representations. Unfortunately they omit the sequential correlation of attributes and consume large quantity of human labors to label the landmarks. In this paper, we propose a deep multi-task cross-domain hashing termed \textit{DMCH}, in which cross-domain embedding and sequential attribute learning are modeled simultaneously. Sequential attribute learning not only provides the semantic guidance for embedding, but also generates rich attention on discriminative local details (\textit{e.g.,} black buttons) of clothing items without requiring extra landmark labels. This leads to promising performance and 306$\times$ boost on efficiency when compared with the state-of-the-art models, which is demonstrated through rigorous experiments on two public fashion datasets.

cs.CV↗

arXiv2018

Coarse-to-Fine Annotation Enrichment for Semantic Segmentation Learning

Rich high-quality annotated data is critical for semantic segmentation learning, yet acquiring dense and pixel-wise ground-truth is both labor- and time-consuming. Coarse annotations (e.g., scribbles, coarse polygons) offer an economical alternative, with which training phase could hardly generate satisfactory performance unfortunately. In order to generate high-quality annotated data with a low time cost for accurate segmentation, in this paper, we propose a novel annotation enrichment strategy, which expands existing coarse annotations of training data to a finer scale. Extensive experiments on the Cityscapes and PASCAL VOC 2012 benchmarks have shown that the neural networks trained with the enriched annotations from our framework yield a significant improvement over that trained with the original coarse labels. It is highly competitive to the performance obtained by using human annotated dense annotations. The proposed method also outperforms among other state-of-the-art weakly-supervised segmentation methods.

cs.CV↗

arXiv2018

Revisiting non-Gaussianity from non-attractor inflation models

Non-attractor inflation is known as the only single field inflationary scenario that can violate non-Gaussianity consistency relation with the Bunch-Davies vacuum state and generate large local non-Gaussianity. However, it is also known that the non-attractor inflation by itself is incomplete and should be followed by a phase of slow-roll attractor. Moreover, there is a transition process between these two phases. In the past literature, this transition was approximated as instant and the evolution of non-Gaussianity in this phase was not fully studied. In this paper, we follow the detailed evolution of the non-Gaussianity through the transition phase into the slow-roll attractor phase, considering different types of transition. We find that the transition process has important effect on the size of the local non-Gaussianity. We first compute the net contribution of the non-Gaussianities at the end of inflation in canonical non-attractor models. If the curvature perturbations keep evolving during the transition - such as in the case of smooth transition or some sharp transition scenarios - the $\mathcal{O}(1)$ local non-Gaussianity generated in the non-attractor phase can be completely erased by the subsequent evolution, although the consistency relation remains violated. In extremal cases of sharp transition where the super-horizon modes freeze immediately right after the end of the non-attractor phase, the original non-attractor result can be recovered. We also study models with non-canonical kinetic terms, and find that the transition can typically contribute a suppression factor in the squeezed bispectrum, but the final local non-Gaussianity can still be made parametrically large.

astro-ph.CO↗

arXiv2017

Strongly Coupled Quasi-Single Field Inflation

We study the power spectrum of quasi-single field inflation where strong coupling is considered. The contribution from the massive propagator can be divided into local and non-local contributions. The local one is the leading contribution and is power-law suppressed as a function of mass, while the non-local contribution is exponentially suppressed in the large mass limit. For the local contribution, it is possible to use the effective field theory approach to study the power spectrum in the strongly coupled region of the parameter space. For the non-local contribution, we developed a partial effective field theory method to simplify the calculation: When there are multiple massive propagators, one can fully compute it after integrating out all but one massive propagator by effective field theory. The result retains the "standard clock" signal, which is interesting for probing the expansion history of the primordial universe and the physics of a "cosmological collider". The error involved compared to the full calculation is power law suppressed by the effective mass of the heavy field.

hep-th↗

arXiv2016

Trajectory Prediction of Rotating Objects in Viscous Fluid: Based on Kinematic Investigation of Magnus Glider

The case of a rotating object traveling through viscous fluid appears in many phenomena like the banana ball and missile movement. In this work, we build a model to predict the trajectory of such rotating objects with near-cylinder geometry. The analytical expression of Magnus force is given and a wind tunnel experiment is carried out, which shows the Magnus force is well proportional to the product of angular velocity and centroid velocity. The trajectory prediction is consistent with the trajectory record experiment of Magnus glider, which implies the validity and robustness of this model.

physics.flu-dyn↗

arXiv2016

Features from the non-attractor beginning of inflation

We study the effects of the non-attractor initial conditions for the canonical single-field inflation. The non-attractor stage can last only several $e$-folding numbers, and should be followed by hilltop inflation. This two-stage evolution leads to large scale suppression in the primordial power spectrum, which is favored by recent observations. Moreover we give a detailed calculation of primordial non-Guassianity due to the "from non-attractor to slow-roll" transition, and find step features in the local and equilateral shapes. We conclude that a plateau-like inflaton potential with an initial non-attractor phase yields interesting features in both power spectrum and bispectrum.

astro-ph.CO↗

arXiv2015

Fluctuations of particle motion in granular avalanches - from the microscopic to the macroscopic scales

In this study, we have investigated the fluctuations of particle motion, i.e. the non-affine motion, during the avalanche process, discovering a rich dynamics from the microscopic to the macroscopic scales. We find that there is strong correlation between the magnitude of the velocity fluctuation and the velocity magnitude in the spatial and temporal domains. The possible connection between this finding and STZ is discussed based on the direct measurement of the T1 events. In addition, the velocity magnitude of the system and the stress fluctuations of the system are strongly correlated temporally. Our finding will pose challenges to the development of more rigorous theories to describe the avalanche dynamics based on the microscopic approach. Moreover, our finding presents a plausible mechanism of the particle entrainment in a simple system.

cond-mat.soft↗