Recently, neural models have been proposed for headline generation by
learning to map documents to headlines with recurrent neural networks.
Nevertheless, as traditional neural network utilizes maximum likelihood
estimation for parameter optimization, it essentially constrains the expected
training objective within word level rather than sentence level. Moreover, the
performance of model prediction significantly relies on training data
distribution. To overcome these drawbacks, we employ minimum risk training
strategy in this paper, which directly optimizes model parameters in sentence
level with respect to evaluation metrics and leads to significant improvements
for headline generation. Experiment results show that our models outperforms
state-of-the-art systems on both English and Chinese headline generation tasks.
One of the most widely used samplers in practice is the component-wise
Metropolis-Hastings (CMH) sampler that updates in turn the components of a
vector valued Markov chain using accept-reject moves generated from a proposal
distribution. When the target distribution of a Markov chain is irregularly
shaped, a `good' proposal distribution for one part of the state space might be
a `poor' one for another part of the state space. We consider a component-wise
multiple-try Metropolis (CMTM) algorithm that can automatically choose from a
set of candidate moves sampled from different distributions. The computational
efficiency is increased using an adaptation rule for the CMTM algorithm that
dynamically builds a better set of proposal distributions as the Markov chain
runs. The ergodicity of the adaptive chain is demonstrated theoretically. The
performance is studied via simulations and real data examples.
Let $M$ be a closed smooth manifold and let $f:M\to M$ be a diffeomorphism.
$C^1$-generically, a continuum-wise expansive satisfies Axiom A without cycles.
Moreover, there is a partially hyperbolic diffeomorphism $f$ such that it is
not continuum-wise expansive.
Authors' comments: 10pages
Using pointwise semigroup techniques, we establish sharp rates of decay in
space and time of a perturbed reaction diffusion front to its time-asymptotic
limit. This recovers results of Sattinger, Henry and others of time-exponential
convergence in weighted $L^p$ and Sobolev norms, while capturing the new
feature of spatial diffusion at Gaussian rate. Novel features of the argument
are a point-wise Green function decomposition reconciling spectral
decomposition and short-time Nash-Aronson estimates and an instantaneous
tracking scheme similar to that used in the study of stability of viscous shock
waves.
Authors' comments: arXiv admin note: substantial text overlap with arXiv:1004.0909 by
other authors
Object proposal is essential for current state-of-the-art object detection
pipelines. However, the existing proposal methods generally fail in producing
results with satisfying localization accuracy. The case is even worse for small
objects which however are quite common in practice. In this paper we propose a
novel Scale-aware Pixel-wise Object Proposal (SPOP) network to tackle the
challenges. The SPOP network can generate proposals with high recall rate and
average best overlap (ABO), even for small objects. In particular, in order to
improve the localization accuracy, a fully convolutional network is employed
which predicts locations of object proposals for each pixel. The produced
ensemble of pixel-wise object proposals enhances the chance of hitting the
object significantly without incurring heavy extra computational cost. To solve
the challenge of localizing objects at small scale, two localization networks
which are specialized for localizing objects with different scales are
introduced, following the divide-and-conquer philosophy. Location outputs of
these two networks are then adaptively combined to generate the final proposals
by a large-/small-size weighting network. Extensive evaluations on PASCAL VOC
2007 show the SPOP network is superior over the state-of-the-art models. The
high-quality proposals from SPOP network also significantly improve the mean
average precision (mAP) of object detection with Fast-RCNN framework. Finally,
the SPOP network (trained on PASCAL VOC) shows great generalization performance
when testing it on ILSVRC 2013 validation set.
Authors' comments: accepted by IEEE Transactions on Image Processing
Motivated by a question of Bumagin and Wise, we construct a continuum of
finitely generated, residually finite groups whose outer automorphism groups
are pairwise non-isomorphic finitely generated, non-recursively-presentable
groups. These are the first examples of such residually finite groups.
Authors' comments: 8 pages
We present a detailed study based on infrared photometry of all Galactic RV
Tauri stars from the General Catalogue of Variable Stars (GCVS). RV Tauri stars
are the brightest among the population II Cepheids. They are thought to evolve
away from the asymptotic giant branch (AGB) towards the white dwarf domain.
IRAS detected several RV Tauri stars because of their large IR excesses and it
was found that they occupy a specific region in the [12] - [25], [25] - [60]
IRAS two-colour diagram. We used the all sky survey of WISE to extend these
studies and compare the infrared properties of all RV Tauri stars in the GCVS
with a selected sample of post-AGB objects with the goal to place the RV Tauri
pulsators in the context of post-AGB evolution. Moreover, we correlated the IR
properties of both the RV Tauri stars and the comparison sample with other
observables like binarity and the presence of a photospheric chemical anomaly
called depletion. We find that Galactic RV Tauri stars display a range of
infrared properties and we differentiate between disc sources, objects with no
IR excess and objects for which the spectral energy distribution (SED) is
uncertain. We obtain a clear correlation between disc sources and binarity. RV
Tauri stars with a variable mean magnitude are exclusively found among the disc
sources. We also find evidence for disc evolution among the binaries.
Furthermore our studies show that the presence of a disc seems to be a
necessary but not sufficient condition for the depletion process to become
efficient.
Authors' comments: 14 pages, 13 figures, 6 tables. Accepted for publication in Monthly
Notices of the Royal Astronomical Society (MNRAS)
We revisit the idea of brain damage, i.e. the pruning of the coefficients of
a neural network, and suggest how brain damage can be modified and used to
speedup convolutional layers. The approach uses the fact that many efficient
implementations reduce generalized convolutions to matrix multiplications. The
suggested brain damage process prunes the convolutional kernel tensor in a
group-wise fashion by adding group-sparsity regularization to the standard
training process. After such group-wise pruning, convolutions can be reduced to
multiplications of thinned dense matrices, which leads to speedup. In the
comparison on AlexNet, the method achieves very competitive performance.
We address the problem of contour detection via per-pixel classifications of
edge point. To facilitate the process, the proposed approach leverages with
DenseNet, an efficient implementation of multiscale convolutional neural
networks (CNNs), to extract an informative feature vector for each pixel and
uses an SVM classifier to accomplish contour detection. In the experiment of
contour detection, we look into the effectiveness of combining per-pixel
features from different CNN layers and verify their performance on BSDS500.
Authors' comments: 2 pages. arXiv admin note: substantial text overlap with
arXiv:1412.6857
We present a morphological study of nebulae around Wolf-Rayet (WR) stars
using archival narrow-band optical and Wide-field Infrared Survey Explorer
(WISE) infrared images. The comparison among WISE images in different bands and
optical images proves to be a very efficient procedure to identify the nebular
emission from WR nebulae, and to disentangle it from that of the ISM material
along the line of sight. In particular, WR nebulae are clearly detected in the
WISE W4 band at 22 $\mu$m. Analysis of available mid-IR Spitzer spectra shows
that the emission in this band is dominated by thermal emission from dust
spatially coincident with the thin nebular shell or most likely with the
leading edge of the nebula. The WR nebulae in our sample present different
morphologies that we classified into well defined WR bubbles (bubble ${\cal
B}$-type nebulae), clumpy and/or disrupted shells (clumpy/disrupted ${\cal
C}$-type nebulae), and material mixed with the diffuse medium (mixed ${\cal
M}$-type nebulae). The variety of morphologies presented by WR nebulae shows a
loose correlation with the central star spectral type, implying that the
nebular and stellar evolutions are not simple and may proceed according to
different sequences and time-lapses. We report the discovery of an obscured
shell around WR35 only detected in the infrared.
Authors' comments: 11 pages, 6 figures, plus 23 appendix figures; to appear in Astronomy
and Astrophysics
We present photometry of images from the Wide-Field Infrared Survey Explorer
(WISE; Wright et al. 2010) of over 400 million sources detected by the Sloan
Digital Sky Survey (SDSS; York et al. 2000). We use a "forced photometry"
technique, using measured SDSS source positions, star-galaxy separation and
galaxy profiles to define the sources whose fluxes are to be measured in the
WISE images. We perform photometry with The Tractor image modeling code,
working on our "unWISE" coaddds and taking account of the WISE point-spread
function and a noise model. The result is a measurement of the flux of each
SDSS source in each WISE band. Many sources have little flux in the WISE bands,
so often the measurements we report are consistent with zero. However, for many
sources we get three- or four-sigma measurements; these sources would not be
reported by the WISE pipeline and will not appear in the WISE catalog, yet they
can be highly informative for some scientific questions. In addition, these
small-signal measurements can be used in stacking analyses at catalog level.
The forced photometry approach has the advantage that we measure a consistent
set of sources between SDSS and WISE, taking advantage of the resolution and
depth of the SDSS images to interpret the WISE images; objects that are
resolved in SDSS but blended together in WISE still have accurate measurements
in our photometry. Our results, and the code used to produce them, are publicly
available at http://unwise.me.
We present 20 WISE-selected galaxies with bolometric luminosities L_bol >
10^14 L_sun, including five with infrared luminosities L_IR = L(rest 8-1000
micron) > 10^14 L_sun. These "extremely luminous infrared galaxies," or ELIRGs,
were discovered using the "W1W2-dropout" selection criteria which requires
marginal or non-detections at 3.4 and 4.6 micron (W1 and W2, respectively) but
strong detections at 12 and 22 micron in the WISE survey. Their spectral energy
distributions are dominated by emission at rest-frame 4-10 micron, suggesting
that hot dust with T_d ~ 450K is responsible for the high luminosities. These
galaxies are likely powered by highly obscured AGNs, and there is no evidence
suggesting these systems are beamed or lensed. We compare this WISE-selected
sample with 116 optically selected quasars that reach the same L_bol level,
corresponding to the most luminous unobscured quasars in the literature. We
find that the rest-frame 5.8 and 7.8 micron luminosities of the WISE-selected
ELIRGs can be 30-80% higher than that of the unobscured quasars. The existence
of AGNs with L_bol > 10^14 L_sun at z > 3 suggests that these supermassive
black holes are born with large mass, or have very rapid mass assembly. For
black hole seed masses ~ 10^3 M_sun, either sustained super-Eddington accretion
is needed, or the radiative efficiency must be <15%, implying a black hole with
slow spin, possibly due to chaotic accretion.
Authors' comments: 17 pages in emulateapj format, including 11 figures and 5 tables. ApJ
in press
The Wide-Field Infrared Survey Explorer (WISE; Wright et al. 2010) satellite
observed the full sky in four mid-infrared bands in the 2.8 to 28 micron range.
The primary mission was completed in 2010. The WISE team have done a superb job
of producing a series of high-quality, well-documented, complete Data Releases
in a timely manner. However, the "Atlas Image" coadds that are part of the
recent AllWISE and previous data releases were intentionally blurred.
Convolving the images by the point-spread function while coadding results in
"matched-filtered" images that are close to optimal for detecting isolated
point sources. But these matched-filtered images are sub-optimal or
inappropriate for other purposes. For example, we are photometering the WISE
images at the locations of sources detected in the Sloan Digital Sky Survey
(York et al. 2000) through forward modeling, and this blurring decreases the
available signal-to-noise by effectively broadening the point-spread function.
This paper presents a new set of coadds of the WISE images that have not been
blurred. These images retain the intrinsic resolution of the data and are
appropriate for photometry preserving the available signal-to-noise. Users
should be cautioned, however, that the W3- and W4-band coadds contain artifacts
around large, bright structures (large galaxies, dusty nebulae, etc);
eliminating these artifacts is the subject of ongoing work. These new coadds,
and the code used to produce them, are publicly available at http://unwise.me .
The Integrated Sachs-Wolfe effect (ISW) measures the decay of the
gravitational potential due to cosmic acceleration and is thus a direct probe
of Dark Energy. In some of the earlier studies, the amplitude of the ISW effect
was found to be in tension with the predictions of the standard $\Lambda$CDM
model. We measure the cross-power of galaxies and AGN from the WISE mission
with CMB temperature data from WMAP9 in order to provide an independent
measurement of the ISW amplitude. Cross-correlations with the recently released
Planck lensing potential maps are used to calibrate the bias and contamination
fraction of the sources, thus avoiding systematic effects that could be present
when using auto-spectra to measure bias. We find an amplitude of the
cross-power of $\mathcal{A} = 1.24\pm 0.47$ from the galaxies and $\mathcal{A}
= 0.88 \pm 0.74$ from the AGN, fully consistent with the $\Lambda$CDM
prediction of $\mathcal{A} =1$. The ISW measurement signal-to-noise ratio is
2.7 and 1.2 respectively, giving a combined significance close to $3 \sigma$.
Comparing the amplitudes of the galaxy and AGN cross-correlations, which arise
from different redshifts, we find no evidence for redshift evolution in Dark
Energy properties, consistent with a Cosmological Constant.
Authors' comments: 9 pages, 8 figures
Using data from the all-sky Wide-Field Infrared Survey Explorer (WISE)
satellite, we made a catalog of over 8000 Galactic HII regions and HII region
candidates by searching for their characteristic mid-infrared (MIR) morphology.
WISE has sufficient sensitivity to detect the MIR emission from HII regions
located anywhere in the Galactic disk. We believe this is the most complete
catalog yet of regions forming massive stars in the Milky Way. Of the ~8000
cataloged sources, ~1500 have measured radio recombination line (RRL) or
H$\alpha$ emission, and are thus known to be HII regions. This sample improves
on previous efforts by resolving HII region complexes into multiple sources and
by removing duplicate entries. There are ~2500 candidate HII regions in the
catalog that are spatially coincident with radio continuum emission. Our
group's previous RRL studies show that ~95% of such targets are HII regions. We
find that ~500 of these candidates are also positionally associated with known
HII region complexes, so the probability of their being bona fide HII regions
is even higher. At the sensitivity limits of existing surveys, ~4000 catalog
sources show no radio continuum emission. Using data from the literature, we
find distances for ~1500 catalog sources, and molecular velocities for ~1500
HII region candidates.
Authors' comments: Accepted to ApJS. Play with catalog contents here:
http://astro.phys.wvu.edu/wise/
With the goal of investigating the degree to which the total infrared
luminosity ($L_{\rm TIR}$) traces the star formation rate (SFR), we analyze the
$L_{\rm TIR}$ from the dust in a sample of $\sim$ 6000 star-forming galaxies,
based on the 3.4, 4.6, 12 and 22 $\mu$m data from the Wide-field Infrared
Survey Explorer (WISE) and u, g, r, i, z band data from SDSS DR9. These
star-forming galaxies are selected by matching the WISE All Sky Catalog with
the star-forming galaxy catalog in SDSS DR9 provided by
JHU/MPA\thanks{$http://www.sdss3.org/dr9/spectro/spectroaccess.php$}. The
values of $L_{\rm TIR}$ and SFR are derived from the project Code Investigating
Galaxy Emission (CIGALE). We study the relationship between the $L_{\rm TIR}$
and SFR. From this study, we derive reference SFR indicators for use in our
analysis. Linear correlations between SFR and the $L_{\rm TIR}$ are found, and
calibrations of SFRs based on it are proposed. The calibration holds for
galaxies with verified observations and agrees well with previous works. The
dispersion in the relation between $L_{\rm TIR}$ and SFR could partly be
explained by the galaxy's properties, such as the 4000 {\AA} break.
Authors' comments: 9 pages, submitted to JKAS. arXiv admin note: text overlap with
arXiv:1111.3814, arXiv:1102.1571 by other authors
We make a detailed pulse-wise study of gamma-ray bursts (GRBs) with known
redshift detected by \emph{Fermi}/Gamma Ray Burst Monitor (GBM). The sample
contains 19 GRBs with 43 pulses. We find that the average peak energy is
correlated to the radiated energy (the Amati relation) for individual pulses
with a correlation coefficient of 0.86, which is slightly better than the
correlation for the full GRBs. As the present correlation holds within GRBs, it
is a strong evidence supporting the reliability of such a correlation. We
investigate several aspects of this correlation. (i) We divide our sample into
redshift bins and study the evolution of the correlation. Though there is a
marginal indication of evolution of the correlation, we can conclude that the
present data is consistent with no evolution. (ii) We compare the correlation
in the first or single pulses of these GRBs to that of the rest of the pulses,
and confirm that the correlation is unaffected by the fact that first/single
pulses are generally harder than the rest. Finally, we conclude that the
pulse-wise Amati correlation is more robust and it has the potential of
refining the correlation so that GRB study could be used as a cosmological
tool.
Authors' comments: 7 pages, 6 figures, MNRAS accepted
Mateos et al. (2012) presented a highly reliable and efficient mid-infrared
(MIR) colour-based selection technique for luminous active galactic nuclei
(AGN) using the Wide-field Infrared Survey Explorer (WISE) survey. Here we
evaluate the effectiveness of this technique to identify obscured AGN missed in
X-ray surveys. To do so we study the WISE properties of AGN independently
selected in hard X-ray and optical surveys. We use the largest catalogue of 887
[O III]{\lambda}5007-selected type 2 quasars (QSO2s) at z<0.83 in the
literature from the Sloan Digital Sky Survey, and the 258 hard (>4.5 keV)
X-ray-selected AGN from the Bright Ultrahard XMM-Newton Survey (BUXS). The
fraction of SDSS QSO2s in our infrared AGN selection region (wedge) increases
with the AGN luminosity, reaching 66.1+4.5_4.7% at the highest [O III]
luminosities in the sample. This fraction is substantially lower than for the
BUXS type 1 AGN (96.1+3.0_6.3%), but consistent, within the uncertainties, with
that for the BUXS type 2 AGN (75.0+14.1_19.1%) with the same luminosity. The
SDSS QSO2s appear to reside in more luminous (massive) hosts than the BUXS AGN,
due to the tight magnitude limits applied in the SDSS spectroscopic target
selection. Since host galaxy dilution can reduce substantially the
effectiveness of MIR-based techniques, this may explain the lower fraction of
SDSS QSO2s in the WISE AGN wedge. The fraction of SDSS QSO2s identified as
Compton-thick candidates that fall in the wedge is consistent with the fraction
of all SDSS QSO2s in that zone. At the AGN luminosities involved in the
comparison, Compton-thick and Compton-thin SDSS QSO2s have similar WISE colour
distributions. We conclude that at high luminosities and z<1 our MIR technique
is very effective at identifying both Compton-thin and Compton-thick AGN.
Authors' comments: 16 pages, 10 figures, accepted for publication in MNRAS
Linear NDCG is used for measuring the performance of the Web content quality
assessment in ECML/PKDD Discovery Challenge 2010. In this paper, we will prove
that the DCG error equals a new pair-wise loss.
Authors' comments: 5 pages, 3 figures
When using deep, multi-layered architectures to build generative models of
data, it is difficult to train all layers at once. We propose a layer-wise
training procedure admitting a performance guarantee compared to the global
optimum. It is based on an optimistic proxy of future performance, the best
latent marginal. We interpret auto-encoders in this setting as generative
models, by showing that they train a lower bound of this criterion. We test the
new learning procedure against a state of the art method (stacked RBMs), and
find it to improve performance. Both theory and experiments highlight the
importance, when training deep architectures, of using an inference model (from
data to hidden variables) richer than the generative model (from hidden
variables to data).
