arXiv subjects

Sungeun Kim

Publications and source records attributed to Sungeun Kim.

18 recordsLinked to original sources

arXiv2023

Robust Asymmetric Loss for Multi-Label Long-Tailed Learning

In real medical data, training samples typically show long-tailed distributions with multiple labels. Class distribution of the medical data has a long-tailed shape, in which the incidence of different diseases is quite varied, and at the same time, it is not unusual for images taken from symptomatic patients to be multi-label diseases. Therefore, in this paper, we concurrently address these two issues by putting forth a robust asymmetric loss on the polynomial function. Since our loss tackles both long-tailed and multi-label classification problems simultaneously, it leads to a complex design of the loss function with a large number of hyper-parameters. Although a model can be highly fine-tuned due to a large number of hyper-parameters, it is difficult to optimize all hyper-parameters at the same time, and there might be a risk of overfitting a model. Therefore, we regularize the loss function using the Hill loss approach, which is beneficial to be less sensitive against the numerous hyper-parameters so that it reduces the risk of overfitting the model. For this reason, the proposed loss is a generic method that can be applied to most medical image classification tasks and does not make the training process more time-consuming. We demonstrate that the proposed robust asymmetric loss performs favorably against the long-tailed with multi-label medical image classification in addition to the various long-tailed single-label datasets. Notably, our method achieves Top-5 results on the CXR-LT dataset of the ICCV CVAMD 2023 competition. We opensource our implementation of the robust asymmetric loss in the public repository: https://github.com/kalelpark/RAL.

cs.CV↗

arXiv2020

An odd $[1,b]$-factor in regular graphs from eigenvalues

An odd $[1,b]$-factor of a graph $G$ is a spanning subgraph $H$ such that for each vertex $v \in V(G)$, $d_H(v)$ is odd and $1\le d_H(v) \le b$. Let $λ_3(G)$ be the third largest eigenvalue of the adjacency matrix of $G$. For positive integers $r \ge 3$ and even $n$, Lu, Wu, and Yang [10] proved a lower bound for $λ_3(G)$ in an $n$-vertex $r$-regular graph $G$ to gurantee the existence of an odd $[1,b]$-factor in $G$. In this paper, we improve the bound; it is sharp for every $r$.

math.CO↗

arXiv2019

CO Outflow Survey of 68 Very Low Luminosity Objects: a Search for Proto-Brown Dwarf Candidates

We present the results of a systematic search for molecular outflows in 68 Very Low Luminosity Objects (VeLLOs) from single-dish observations in CO isotopologues which find 16 VeLLOs showing clear outflow signatures in the CO maps. With additional three VeLLOs from the literature, we analyzed the outflow properties for 19 VeLLOs, identifying 15 VeLLOs as proto-Brown Dwarf (BD) candidates and four VeLLOs as likely faint protostar candidates. The proto-BD candidates are found to have a mass accretion rate ($\sim 10^{-8} - 10^{-7}$ $\rm M_{\odot}$ yr$^{-1}$) lower than that of the protostar candidates ($\gtrsim 10^{-6}$ $\rm M_{\odot}$ yr$^{-1}$). Their accretion luminosities are similar to or smaller than their internal luminosities, implying that many proto-BD candidates might have had either small accretion activity in a quiescent manner throughout their lifetime, or be currently exhibiting a relatively higher (or episodic) mass accretion than the past. There are strong trends that outflows of many proto-BDs are less active if they are fainter or have less massive envelopes. The outflow forces and internal luminosities for more than half of the proto-BD candidates seem to follow an evolutionary track of a protostar with its initial envelope mass of $\sim$0.08 $\rm M_{\odot}$, indicating that some BDs may form in less massive dense cores in a way similar to normal stars. But, because there also exists a significant fraction (about 40%) of proto-BDs with much weaker outflow force than expected by the relations for protostars, we should not rule out the possibility of other formation mechanism for the BDs.

astro-ph.GA↗

arXiv2013

Global Spectral Energy Distributions of the Large Magellanic Cloud with Interstellar Dust

The effects of dust on infrared emission vary among galaxies of different morphological types. We investigated integrated spectral energy distributions in infrared and submillimeter emissions from the Large Magellanic Cloud based on the observations using the Herschel Space Observatory and near- to mid-infrared observations from the Spitzer Space Telescope. We also used IRAS and WMAP observations to constrain the SEDs. We explain the observations by using dust models with different grain size distributions in the interstellar medium of the LMC, noting that the LMC has undergone processes that are different from those in the Milky Way. We determined a spectral index and a normalization factor in the range of -3.5 to -3.45 with grain radii in the range of 1 nm-300 nm for the silicate grain and 2 nm-1 um for the graphite grain. The best fit to the observed SED was obtained with a spectral index of -3.47. The grain size distribution is described using a power law but with a break that is introduced below a_b, where a larger exponent is used. Changing graphite grain size distribution significantly changed the SED pattern within the observational uncertainties. Based on the SED fits to the observations from submillimeter wavelengths to the infrared radiation in the LMC using GRASIL (Silva et al. 1998), we obtained a reasonable set of parameter values in chemical and geometric space together with the grain size distributions of a modified MRN model with the LMC extinction curve.

astro-ph.CO↗

arXiv2013

Submm/mm Galaxy Counterpart Identification Using a Characteristic Density Distribution

We present a new submm/mm galaxy counterpart identification technique which builds on the use of Spitzer IRAC colors as discriminators between likely counterparts and the general IRAC galaxy population. Using 102 radio- and SMA-confirmed counterparts to AzTEC sources across three fields (GOODS-N, GOODS-S, and COSMOS), we develop a non-parametric IRAC color-color characteristic density distribution (CDD), which, when combined with positional uncertainty information via likelihood ratios, allows us to rank all potential IRAC counterparts around SMGs and calculate the significance of each ranking via the reliability factor. We report all robust and tentative radio counterparts to SMGs, the first such list available for AzTEC/COSMOS, as well as the highest ranked IRAC counterparts for all AzTEC SMGs in these fields as determined by our technique. We demonstrate that the technique is free of radio bias and thus applicable regardless of radio detections. For observations made with a moderate beamsize (~18"), this technique identifies ~85 per cent of SMG counterparts. For much larger beamsizes (>30"), we report identification rates of 33-49 per cent. Using simulations, we demonstrate that this technique is an improvement over using positional information alone for observations with facilities such as AzTEC on the LMT and SCUBA-2 on JCMT.

astro-ph.IM↗

arXiv2012

Power Spectrum Analysis of the OMC1 Image at 1.1 mm Wavelength

We present a 1.1mm emission map of the OMC1 region observed with AzTEC, a new large-format array composed of 144 silicon-nitride micromesh bolometers, that was in use at the James Clerk Maxwell Telescope (JCMT). These AzTEC observations reveal dozens of cloud cores and a tail of filaments in a manner that is almost identical to the submillimeter continuum emission of the entire OMC1 region at 450 and 850 micronm. We perform Fourier analysis of the image with a modified periodogram and the density power spectrum, which provides the distribution of the length scale of the structures, is determined. The expected value of the periodogram converges to the resulting power spectrum in the mean squared sense. The present analysis reveals that the power spectrum steepens at relatively smaller scales. At larger scales, the spectrum flattens and the power law becomes shallower. The power spectra of the 1.1mm emission show clear deviations from a single power law. We find that at least three components of power law might be fitted to the calculated power spectrum of the 1.1mm emission. The slope of the best fit power law, γ~ -2.7 is similar to those values found in numerical simulations. The effects of beam size and the noise spectrum on the shape and slope of the power spectrum are also included in the present analysis. The slope of the power law changes significantly at higher spatial frequency as the beam size increases.

astro-ph.GA↗

arXiv2012

Structure Analysis of OMC1 at 1.1 mm

We present a 1.1 mm emission map of the OMC1 region observed with AzTEC, a new large-format array composed of 144 silicon-nitride micromesh bolometers that was in use at the James Clerk Maxwell Telescope (JCMT). The AzTEC observations of the OMC1 region at 1.1 mm reveal dozens of cloud cores and a tail of filaments in a manner that is almost identical to the submillimeter continuum emission of the entire OMC1 region at 450 and 850 micronm. The density power spectrum provides the size distribution of the structures. We find that a single power law might be fitted to the calculated power spectrum of the 1.1 mm emission between 0.3 pc and 0.03 pc. The slope of the best fit power law is γ\sim -2.6 and is similar to the spectral index of the power spectrum of γ\sim -2.7 found in numerical simulations. However, there is a distinct spectral break in the power spectrum at a characteristic scale of ~0.3 pc in OMC1. The effects of beam size and noise spectrum on the shape and slope of the power spectrum are also included in the present analysis. The slope of the power law and a range of different scales change at scales below ~0.3 pc as the beam size increases.

astro-ph.GA↗

arXiv2010

Dense Clumps in Giant Molecular Clouds in the Large Magellanic Cloud: Density and Temperature Derived from $^{13}$CO($J=3-2$) Observations

In order to precisely determine temperature and density of molecular gas in the Large Magellanic Cloud, we made observations of optically thin $^{13}$CO($J=3-2$) transition by using the ASTE 10m telescope toward 9 peaks where $^{12}$CO($J=3-2$) clumps were previously detected with the same telescope. The molecular clumps include those in giant molecular cloud (GMC) Types I (with no signs of massive star formation), II (with HII regions only), and III (with HII regions and young star clusters). We detected $^{13}$CO($J=3-2$) emission toward all the peaks and found that their intensities are 3 -- 12 times lower than those of $^{12}$CO($J=3-2$). We determined the intensity ratios of $^{12}$CO($J=3-2$) to $^{13}$CO($J=3-2$), $R^{12/13}_{3-2}$, and $^{13}$CO($J=3-2$) to $^{13}$CO($J=1-0$), $R^{13}_{3-2/1-0}$, at 45$\arcsec$ resolution. These ratios were used for radiative transfer calculations in order to estimate temperature and density of the clumps. The parameters of these clumps range kinetic temperature $T\mathrm{_{kin}}$ = 15 -- 200 K, and molecular hydrogen gas density $n(\mathrm{H_2})$ = 8$\times 10^2$ -- 7$\times 10^3$ cm$^{-3}$. We confirmed that the higher density clumps show higher kinetic temperature and that the lower density clumps lower kinetic temperature at a better accuracy than in the previous work. The kinetic temperature and density increase generally from a Type I GMC to a Type III GMC. We interpret that this difference reflects an evolutionary trend of star formation in molecular clumps. The $R^{13}_{3-2/1-0}$ and kinetic temperature of the clumps are well correlated with H$α$ flux, suggesting that the heating of molecular gas $n(\mathrm{H_2})$ = $10^3$ -- $10^4$ cm$^{-3}$ can be explained by stellar FUV photons.

astro-ph.GA↗

arXiv2010

Dust/gas correlations from Herschel Observations

Spitzer and IRAS observations of the LMC suggest an excess of FIR emission with respect to the gas surface density traced by 12CO and HI 21 cm emission lines. This "FIR excess" is noticeable near molecular clouds in the LMC, and has usually been interpreted as the presence of a self-shielded H2 component not traced by CO molecular clouds' envelopes. Based on Herschel observations, we examine the correlation between gas and dust at higher resolution than previously achieved. We consider three additional causes for the FIR excess: X factor, FIR dust emissivity, and gas-to-dust ratio variations between the diffuse and dense phases of the ISM. We examine the structure of NT80 and NT71, two molecular clouds detected in the NANTEN 12CO survey of the LMC. Dust surface density maps are derived from the HERITAGE data. The gas phase is traced by MAGMA 12CO and ATCA HI 21 cm observations of the LMC. The dust emissivity, gas-to-dust ratio, and X factor required to match the dust and gas surface densities are derived, and their correlations with the dust surface density are examined. The dust surface density is spatially correlated with the atomic and molecular gas phases. The dust temperature is consistently lower in the dense phase of the ISM than in the diffuse phase. We confirm variations in the ratio of FIR emission to gas surface density derived from HI and CO observations. There is an excess of FIR emission, spatially correlated with regions of intermediate HI and dust surface densities (Av = 1-2), and little or no CO. While there is no significant trend in the dust emissivity or gas-to-dust ratio with dust surface density, the X factor is enhanced at Av = 1-2. We conclude that H2 envelopes not traced by CO and X factor variations close to the CO boundary are more likely to cause these deviations between FIR emission and gas surface density than gas-to-dust ratio or emissivity variations.

astro-ph.GA↗

arXiv2007

Evidence for a Population of High-Redshift Submillimeter Galaxies from Interferometric Imaging

We have used the Submillimeter Array to image a flux limited sample of seven submillimeter galaxies, selected by the AzTEC camera on the JCMT at 1.1 mm, in the COSMOS field at 890um with 2" resolution. All of the sources - two radio-bright and five radio-dim - are detected as single point-sources at high significance (> 6σ), with positions accurate to 0.2" that enable counterpart identification at other wavelengths observed with similarly high angular resolution. All seven have IRAC counterparts, but only two have secure counterparts in deep HST/ACS imaging. As compared to the two radio-bright sources in the sample, and those in previous studies, the five radio-dim sources in the sample (1) have systematically higher submillimeter-to-radio flux ratios, (2) have lower IRAC 3.6-8.0um fluxes, and (3) are not detected at 24um. These properties, combined with size constraints at 890um (θ< 1.2"), suggest that the radio-dim submillimeter galaxies represent a population of very dusty starbursts, with physical scales similar to local ultraluminous infrared galaxies, and an average redshift higher than radio-bright sources.

astro-ph↗

arXiv2004

AST/RO Observations of CO J=4-3 Emission from the N44 Complex in the Large Magellanic Cloud

We present Antarctic Submillimeter Telescope and Remote Observatory (AST/RO) observations of 12CO J=4-3 and C I emission in the N44 H II complex in the Large Magellanic Cloud. We detected strong 12CO J=4-3 emission toward the H II region called as N44BC, which is located on the rim of an expanding giant shell in the N44 region. Analysis with a photodissociation region (PDR) model showed that the 12CO J=4-3 emitting cloud is very dense, with n ~ 10^5 cm^-3. We also note that there is a high-velocity component associated with the 12CO J=4-3 emission. This probably originates from molecular material accelerated as a result of the motion induced by the expanding giant shell surrounding LH47 in the N44 complex. We found that the kinetic energy of this high-velocity gas observed in the CO J=4-3 emission toward the rim of the expanding H II shell is at least an order of magnitude higher than the kinetic energy derived for the H I and H II gas in this region.

astro-ph↗

arXiv2002

AST/RO Observations of CO J=7-6 and J=4-3 Emission toward the Galactic Center Region

We present position-velocity strip maps of the Galactic Center region in the CO J=7-6 and J=4-3 transitions observed with the Antarctic Submillimeter Telescope and Remote Observatory (AST/RO) located at Amundsen-Scott South Pole Station. Emission from the two rotational transitions of 12^CO was mapped at b=0 deg for 3.5 deg > l > -1.5 deg, on a 1' grid with a FWHM beamsize of 58'' at 806 GHz and 105'' at 461 GHz. Previous observations of CO J=4-3 (Martin et al., in preparation) and CI (Ojha et al. 2001) emission from this region show that these lines are distributed in a manner similar to CO J=1-0 (Stark et al. 1987); the (CO J=4-3)/(CO J=1-0) line ratio map is almost featureless across the entire Galactic Center region. In contrast, the CO J=7-6 emission from the Galactic Center is strongly peaked toward the Sgr A and Sgr B molecular complexes. A Large Velocity Gradient (LVG) analysis shows that aside from the two special regions Sgr A and Sgr B, the photon-dominated regions within a few hundred parsecs of the Galactic Center are remarkably uniform in mean density and kinetic temperature at n = 2500 to 4000 cm^{-3} and T=30 to 45K. The (CO J=7-6)/(CO J=4-3) line temperature ratios near Sgr B are a factor of two higher than those observed in the nuclear region of the starburst galaxy M82 (Mao et al. 2000), while the CO(J=7-6)/CO(J=4-3) line temperature ratios around Sgr A are similar to M82. The line ratio on large scales from the Galactic Center region is an order of magnitude less than that from M82.

astro-ph↗

arXiv2001

A New Method to Measure and Map the Gas Scale-Height of Disk Galaxies

We propose a new method to measure and map the gas scale height of nearby disk galaxies. This method is applied successfully to the Australia Telescope Compact Array interferometric HI survey of the Large Magellanic Cloud (LMC); it could also be applied to a significant number of nearby disk galaxies, thanks to the next generation of interferometric facilities, such as the extended VLA and CARMA. The method consists of computing the Spectral Correlation Function (SCF) for a spectral-line map of a face-on galaxy. The SCF quantifies the correlation between spectra at different map positions as a function of their separation, and is sensitive to the properties of both the gas mass distribution and the gas velocity field. It is likely that spatial correlation properties of the gas density and velocity fields in a galactic disk are sensitive to the value of the scale height of the gas disk. A scale-free turbulent cascade is unlikely to extend to scales much larger than the disk scale height, as the disk dynamics on those larger scales should be dominated by two dimensional motions. We find a clear feature in the SCF of the LMC HI disk, on the scale of approximately 180 pc, which we identify as the disk scale height. We are also tentatively able to map variations of the scale height over the disk.

astro-ph↗

arXiv2001

A Comprehensive Look at LH72 in the Context of Supergiant Shell LMC-4

Stellar spectroscopy, UBV photometry, H$α$ imaging, and analysis of data from the ATCA \ion{H}{1} survey of the LMC are combined in a study of the LMC OB association LH 72 and its surroundings. LH 72 lies on the rim of a previously identified \ion{H}{1} shell, SGS-14, and in the interior of LMC-4, one of the LMC's largest known supergiant shells. Our analysis of the \ion{H}{1} data finds that SGS-14 is expanding with velocity $v_{exp}\sim15$ km s$^{-1}$, giving it an expansion age of $\sim$15 Myr. Through the stellar spectroscopy and photometry, we find similar ages for the oldest stars of LH 72, $\sim15-$30 Myr. We confirm that LH 72 contains an age spread of $\sim15-$30 Myr, similar to the range in ages of stars derived for the entire surrounding supergiant shell. Combining analysis of the O and B stars with H$α$ imaging of the \ion{H}{2} region DEM 228, we find that DEM 228 accounts for only 60% of the available ionizing Lyman continuum photons. Comparing the distribution of ionized gas with that of the \ion{H}{1}, we find that DEM 228 and LH 72 are offset by $\sim1-2\arcmin$ from the peak 21-cm emission, towards the interior of SGS-14. Taken together, these results imply that SGS-14 has cleared its interior of gas and triggered the formation of LH 72. On the basis of our results, we suggest that LMC-4 was not formed as unit but by overlapping shells such as SGS-14, and that LH 72 will evolve to produce a stellar arc similar to others seen within LMC-4.

astro-ph↗

arXiv2000

A Fractal Analysis of the HI Emission from the Large Magellanic Cloud

A composite map of HI in the LMC using the ATCA interferometer and the Parkes multibeam telescope was analyzed in several ways in an attempt to characterize the structure of the neutral gas and to find an origin for it. Fourier transform power spectra in 1D, 2D, and in the azimuthal direction were found to be approximate power laws over 2 decades in length. Delta-variance methods also showed the same power-law structure. Detailed models of these data were made using line-of-sight integrals over fractals that are analogous to those generated by simulations of turbulence with and without phase transitions. The results suggested a way to measure directly for the first time the line-of-sight thickness of the cool component of the HI disk of a nearly face-on galaxy. The signature of this thickness was found to be present in all of the measured power spectra. The character of the HI structure in the LMC was also viewed by comparing positive and negative images of the integrated emission. The geometric structure of the high-emission regions was found to be filamentary, whereas the geometric structure of the low-emission (intercloud) regions was found to be patchy and round. This result suggests that compressive events formed the high-emission regions, and expansion events, whether from explosions or turbulence, formed the low-emission regions. The character of the structure was also investigated as a function of scale using unsharp masks. All of these results suggest that most of the ISM in the LMC is fractal, presumably the result of pervasive turbulence, self-gravity, and self-similar stirring.

astro-ph↗

arXiv2000

Carbon in the N159/N160 Complex of the Large Magellanic Cloud

We present a study of carbon in N159/N160, an HII region complex in the low metallicity Large Magellanic Cloud. We have mapped this region, which comprises four distinct molecular clouds spanning a wide range of star-formation activity, in four transitions: 13CO (1-0) CO (2-1) and (4-3), and [CI] (1-0). Combining these data with existing [CII] observations provides a complete picture of the predominant forms of carbon in the gas phase of the ISM. The new CO (2-1) data show that the complex is immersed in an envelope of extended, low level emission, undetected by previous (1-0) mapping efforts. The CO (2-1)/(1-0) ratio in this envelope is >~3, a value consistent with optically thin CO emission. The envelope is also relatively bright in [CI] and [CII], and calculations show that it is mostly photodissociated: it appears to be translucent (Av<1). Neutral carbon emission in the complex unexpectedly peaks at the quiescent southern cloud (N159S). In the northern portion of the map (the N160 nebula), the HII regions prominent in [CII] correspond to holes in the [CI] distribution. Overall we find that, while the [CII]/CO ratio is enhanced with respect to similar complexes in the Milky Way, the [CI]/CO ratio appears to be similar or reduced.

astro-ph↗

arXiv2000

Formation of Cavities, Filaments, and Clumps by the Non-linear Development of Thermal and Gravitational Instabilities in the Interstellar Medium under Stellar Feedback

Based on our high resolution, two-dimensional hydrodynamical simulations, we propose that large cavities may be formed by the nonlinear development of the combined thermal and gravitational instabilities, without need for stellar energy injection in a galaxy modeling the Large Magellanic Clouds (LMC). Our numerical model of the star formation allows us to follow the evolution of the blastwaves due to supernovae in the inhomogeous, multi-phase, and turbulent-like media self-consistently. Formation of kpc-scale inhomogeneity, such as cavities, observed HI map of the LMC, is suppressed by frequent supernovae (average supernova rate for the whole disk is ~0.001/yr). However the supernova explosions are necessary for the hot component (T_g > 10^{6-7} K). Position-velocity maps show that kpc-scale shells/arcs formed through the nonlinear evolution in a model without stellar energy feedback has similar kinematics to explosional phenomena, such as supernovae. We also find that dense clumps and filamentary structure are formed due to a natural consequence of the non-linear evolution of the multi-phage ISM. Although the ISM in a small scale looks turbulent-like and transient, the global structure of the ISM is quasi-stable. We compare the observations of HI and molecular gas of the LMC with the numerically obtained HI and CO map. The morphology and statistical properties of the numerical HI and CO maps are discussed.

astro-ph↗

arXiv1998

The Multi-Phase Medium in the Interstellar Complex N44

We have obtained high-resolution HI observations of N44, one of the largest HII complexes in the Large Magellanic Cloud. The distribution and internal motions of the HI gas show dynamic effects of fast stellar winds and supernova blasts. Numerous HI holes are detected, with the most prominent two corresponding to the optically identified superbubbles Shell 1 and Shell 2. The HI gas associated with Shell 1 shows an expansion pattern similar to that of the ionized gas shell, but the mass and kinetic energy of the HI shell are 3--7 times those of the ionized gas shell. The total kinetic energy of the neutral and ionized gas of Shell 1 is still more than a factor of 5 lower than expected in a pressure-driven superbubble. It is possible that the central OB association was formed in a molecular cloud and a visible superbubble was not fully developed until the ambient molecular gas had been dissociated and cleared away. This hypothesis is supported by the existence of a molecular cloud toward N44 and the fact that the apparent dynamic age of the superbubble Shell 1 is much shorter than the age of its OB association LH47. Accelerated HI gas is detected at the supernova remnant 0523-679. The mass and kinetic energy in the associated HI gas are also much higher than those in the ionized gas of 0523-679. Studies of interstellar gas dynamics using ionized gas alone are clearly inadequate; neutral gas components must be included.

astro-ph↗