arXiv Science⌕ Search

arXiv subjects

Yin Li

Publications and source records attributed to Yin Li.

191 records · Page 11Linked to original sources
arXiv2016

Learning Deep Structure-Preserving Image-Text Embeddings

This paper proposes a method for learning joint embeddings of images and text using a two-branch neural network with multiple layers of linear projections followed by nonlinearities. The network is trained using a large margin objective that combines cross-view ranking constraints with within-view neighborhood structure preservation constraints inspired by metric learning literature. Extensive experiments show that our approach gains significant improvements in accuracy for image-to-text and text-to-image retrieval. Our method achieves new state-of-the-art results on the Flickr30K and MSCOCO image-sentence datasets and shows promise on the new task of phrase localization on the Flickr30K Entities dataset.

cs.CV↗
arXiv2016

Unsupervised Learning of Edges

Data-driven approaches for edge detection have proven effective and achieve top results on modern benchmarks. However, all current data-driven edge detectors require manual supervision for training in the form of hand-labeled region segments or object boundaries. Specifically, human annotators mark semantically meaningful edges which are subsequently used for training. Is this form of strong, high-level supervision actually necessary to learn to accurately detect edges? In this work we present a simple yet effective approach for training edge detectors without human supervision. To this end we utilize motion, and more specifically, the only input to our method is noisy semi-dense matches between frames. We begin with only a rudimentary knowledge of edges (in the form of image gradients), and alternate between improving motion estimation and edge detection in turn. Using a large corpus of video data, we show that edge detectors trained using our unsupervised scheme approach the performance of the same methods trained with full supervision (within 3-5%). Finally, we show that when using a deep network for the edge detector, our approach provides a novel pre-training scheme for object detection.

cs.CV↗
arXiv2016

Separate Universe Consistency Relation and Calibration of Halo Bias

Linear halo bias is the response of dark matter halo number density to a long wavelength fluctuation in the dark matter density. Using abundance matching between separate universe simulations which absorb the latter into a change in the background, we test the consistency relation between the change in a one point function, the halo mass function, and a two point function, the halo-matter cross correlation in the long wavelength limit. We find excellent agreement between the two at the $1-2\%$ level for average halo biases between $1 \lesssim \bar b_1 \lesssim 4$ and no statistically significant deviations at the $4-5\%$ level out to $\bar b_1 \approx 8$. Halo bias inferred assuming instead a universal mass function is significantly different and inaccurate at the 10\% level or more. The separate universe technique provides a way of calibrating linear halo bias efficiently for even highly biased rare halos in the $Λ$CDM model. Observational violation of the consistency relation would indicate new physics, e.g.~in the dark matter, dark energy or primordial non-Gaussianity sectors.

astro-ph.CO↗
arXiv2015

Position Reconstruction of Bubble Formation in Liquid Nitrogen using Piezoelectric Sensors

Cryogenic liquids, particularly liquid xenon and argon, are of interest as detector media for experiments in nuclear and particle physics. Here we present a new detector diagnostic technique using piezoelectric sensors to detect bubbling of the liquid. Bubbling can indicate locations of excess heat dissipation e.g., in immersed electronics. They can also interfere with normal event evolution by scattering of light or by interrupting the drift of ionization charge. In our test apparatus, four sensors are placed in the vacuum space of a double-walled dewar of liquid nitrogen and used to detect and locate a source of bubbling inside the liquid volume. Utilizing the differences in transmitted frequencies through the different media present in the experiment, we find that sound traveling in a direct path from the source to the sensor can be isolated with appropriate filtering. The location of the source is then reconstructed using the time difference of arrivals (TDOA) information. The reconstruction algorithm is shown to have a 95.8% convergence rate and reconstructed positions are self-consistent to an average +/-0.5cm around the mean in x, y, and z. Systematic effects are observed to cause errors in reconstruction when bubbles occur very close to the surfaces of the liquid volume.

physics.ins-det↗
arXiv2014

Super-Sample Signal

When extracting cosmological information from power spectrum measurements, we must consider the impact of super-sample density fluctuations whose wavelengths are larger than the survey scale. These modes contribute to the mean density fluctuation $δ_b$ in the survey and change the power spectrum in the same way as a change in the cosmological background. They can be simply included in cosmological parameter estimation and forecasts by treating $δ_b$ as an additional cosmological parameter enabling efficient exploration of its impact. To test this approach, we consider here an idealized measurement of the matter power spectrum itself in the $Λ$CDM cosmology though our techniques can readily be extended to more observationally relevant statistics or other parameter spaces. Using sub-volumes of large-volume $N$-body simulations for power spectra measured with respect to either the global or local mean density, we verify that the minimum variance estimator of $δ_b$ is both unbiased and has the predicted variance. Parameter degeneracies arise since the response of the matter power spectrum to $δ_b$ and cosmological parameters share similar properties in changing the growth of structure and dilating the scale of features especially in the local case. For matter power spectrum measurements, these degeneracies can lead in certain cases to substantial error degradation and motivates future studies of specific cosmological observables such as galaxy clustering and weak lensing statistics with these techniques.

astro-ph.CO↗
arXiv2014

The Secrets of Salient Object Segmentation

In this paper we provide an extensive evaluation of fixation prediction and salient object segmentation algorithms as well as statistics of major datasets. Our analysis identifies serious design flaws of existing salient object benchmarks, called the dataset design bias, by over emphasizing the stereotypical concepts of saliency. The dataset design bias does not only create the discomforting disconnection between fixations and salient object segmentation, but also misleads the algorithm designing. Based on our analysis, we propose a new high quality dataset that offers both fixation and salient object segmentation ground-truth. With fixations and salient object being presented simultaneously, we are able to bridge the gap between fixations and salient objects, and propose a novel method for salient object segmentation. Finally, we report significant benchmark progress on three existing datasets of segmenting salient objects

cs.CV↗
arXiv2014

Super-Sample Covariance in Simulations

Using separate universe simulations, we accurately quantify super-sample covariance (SSC), the typically dominant sampling error for matter power spectrum estimators in a finite volume, which arises from the presence of super survey modes. By quantifying the power spectrum response to a background mode, this approach automatically captures the separate effects of beat coupling in the quasilinear regime, halo sample variance in the nonlinear regime and a new dilation effect which changes scales in the power spectrum coherently across the survey volume, including the baryon acoustic oscillation scale. It models these effects at typically the few percent level or better with a handful of small volume simulations for any survey geometry compared with directly using many thousands of survey volumes in a suite of large volume simulations. The stochasticity of the response is sufficiently small that in the quasilinear regime, SSC can be alternately included by fitting the mean density in the volume with these fixed templates in parameter estimation. We also test the halo model prescription and find agreement typically at better than the 10% level for the response.

astro-ph.CO↗
arXiv2013

The abundance of voids and the excursion set formalism

We present measurements of the number density of voids in the dark matter distribution from a series of N-body simulations of a ΛCDM cosmology. We define voids as spherical regions of ρ_v = 0.2ρ_m around density minima in order to relate our results to the predicted abundances using the excursion set formalism. Using a linear underdensity of δ_v = -2.7, from a spherical evolution model, we find that a volume conserving model, which does not conserve number density in the mapping from the linear to nonlinear regime, matches the measured abundance to within 16% for a range of void radii 1< r(Mpc/h)<15. This model fixes the volume fraction of the universe which is in voids and assumes that voids of a similar size merge as they expand by a factor of 1.7 to achieve a nonlinear density of ρ_v = 0.2ρ_m today. We find that the model of Sheth & van de Weygaert (2004) for the number density of voids greatly overpredicts the abundances over the same range of scales. We find that the volume conserving model works well at matching the number density of voids measured from the simulations at higher redshifts, z=0.5 and 1, as well as correctly predicting the abundances to within 25% in a simulation of a matter dominated Ω_m = 1 universe. We examine the abundance of voids in the halo distribution and find fewer small, r<10 Mpc/h, voids and many more large, r>10 Mpc/h, voids compared to the dark matter. These results indicate that voids identified in the halo or galaxy distribution are related to the underlying void distribution in the dark matter in a complicated way which merits further study if voids are to be used as a precision probe of cosmology.

astro-ph.CO↗
arXiv2011

The Gauss-Bonnet-Chern Theorem on Riemannian Manifolds

This expository paper contains a detailed introduction to some important works concerning the Gauss-Bonnet-Chern theorem. The study of this theorem has a long history dating back to Gauss's Theorema Egregium (Latin: Remarkable Theorem) and culminated in Chern's groundbreaking work [14] in 1944, which is a deep and wonderful application of Elie Cartan's formalism. The idea and tools in [14] have a great generalization and continue to produce important results till today. In this paper, we give four different proofs of the Gauss-Bonnet-Chern theorem on Riemannian manifolds, namely Chern's simple intrinsic proof, a topological proof, Mathai-Quillen's Thom form proof and McKean-Singer-Patodi's heat equation proof. These proofs are related with remarkable developments in differential geometry such as the Chern-Weil theory, theory of characteristic classes, Mathai-Quillen's formalism and the Atiyah-Singer index theorem. It is through these brilliant achievements the great importance and influence of Chern's insights and ideas are shown. Our purpose here is to use the Gauss- Bonnet-Chern theorem as a guide to expose the reader to some advanced topics in modern differential geometry.

math.DG↗
arXiv2011

Chameleon Halo Modeling in f(R) Gravity

We model the chameleon effect on cosmological statistics for the modified gravity f(R) model of cosmic acceleration. The chameleon effect, required to make the model compatible with local tests of gravity, reduces force enhancement as a function of the depth of the gravitational potential wells of collapsed structure and so is readily incorporated into a halo model by including parameters for the chameleon mass threshold and rapidity of transition. We show that the abundance of halos around the chameleon mass threshold are enhanced by both the merging from below and the lack of merging to larger masses. This property also controls the power spectrum in the nonlinear regime and we provide a description of the transition to the linear regime that is valid for a wide range of f(R) models.

astro-ph.CO↗