Search

Low-rank adaptation (LoRA) has become a standard approach for fine-tuning large foundation models. However, our theoretical understanding of LoRA remains limited as prior analyses of LoRA's training dynamics either rely on linearization arguments or consider highly simplified setups. In this work, we analyze the LoRA loss landscape without such restrictive assumptions. We define two regimes: a ``special regime'', which includes idealized setups where linearization arguments hold, and a ``generic regime'' representing more realistic setups where linearization arguments do not hold. In the generic regime, we show that LoRA training converges to a global minimizer with low rank and small magnitude, or a qualitatively distinct solution with high rank and large magnitude. Finally, we argue that the zero-initialization and weight decay in LoRA training induce an implicit bias toward the low-rank, small-magnitude region of the parameter space -- where global minima lie -- thus shedding light on why LoRA training usually succeeds in finding global minima.

In recent years, masked diffusion models (MDMs) have emerged as a promising alternative approach for generative modeling over discrete domains. Compared to autoregressive models (ARMs), MDMs trade off complexity at training time with flexibility at inference time. At training time, they must learn to solve an exponentially large number of infilling problems, but at inference time, they can decode tokens in essentially arbitrary order. In this work, we closely examine these two competing effects. On the training front, we theoretically and empirically demonstrate that MDMs indeed train on computationally intractable subproblems compared to their autoregressive counterparts. On the inference front, we show that a suitable strategy for adaptively choosing the token decoding order significantly enhances the capabilities of MDMs, allowing them to sidestep hard subproblems. On logic puzzles like Sudoku, we show that adaptive inference can boost solving accuracy in pretrained MDMs from $<7$% to $\approx 90$%, even outperforming ARMs with $7\times$ as many parameters and that were explicitly trained via teacher forcing to learn the right order of decoding.

Thanks to unprecedented language understanding and generation capabilities of large language model (LLM), Retrieval-augmented Code Generation (RaCG) has recently been widely utilized among software developers. While this has increased productivity, there are still frequent instances of incorrect codes being provided. In particular, there are cases where plausible yet incorrect codes are generated for queries from users that cannot be answered with the given queries and API descriptions. This study proposes a task for evaluating answerability, which assesses whether valid answers can be generated based on users' queries and retrieved APIs in RaCG. Additionally, we build a benchmark dataset called Retrieval-augmented Code Generability Evaluation (RaCGEval) to evaluate the performance of models performing this task. Experimental results show that this task remains at a very challenging level, with baseline models exhibiting a low performance of 46.7%. Furthermore, this study discusses methods that could significantly improve performance.

In-context learning (ICL) describes a language model's ability to generate outputs based on a set of input demonstrations and a subsequent query. To understand this remarkable capability, researchers have studied simplified, stylized models. These studies have consistently observed long loss plateaus, during which models exhibit minimal improvement, followed by a sudden, rapid surge of learning. In this work, we reveal that training on multiple diverse ICL tasks simultaneously shortens the loss plateaus, making each task easier to learn. This finding is surprising as it contradicts the natural intuition that the combined complexity of multiple ICL tasks would lengthen the learning process, not shorten it. Our result suggests that the recent success in large-scale training of language models may be attributed not only to the richness of the data at scale but also to the easier optimization (training) induced by the diversity of natural language training data.

Combining Atacama Large Millimeter/sub-millimeter Array CO(2-1) mapping and JWST near- and mid-infrared imaging, we characterize the relationship between CO(2-1) and polycyclic aromatic hydrocarbon (PAH) emission at ~100 pc resolution in 66 nearby star-forming galaxies, expanding the sample size from previous ~100 pc resolution studies by more than an order of magnitude. Focusing on regions of galaxies where most of the gas is likely to be molecular, we find strong correlations between CO(2-1) and 3.3 micron, 7.7 micron, and 11.3 micron PAH emission, estimated from JWST's F335M, F770W, and F1130W filters. We derive power law relations between CO(2-1) and PAH emission, which have indices in the range 0.8-1.2, implying relatively weak variations in the observed CO-to-PAH ratios across the regions that we study. We find that CO-to-PAH ratios and scaling relationships near HII regions are similar to those in diffuse sight lines. The main difference between the two types of regions is that sight lines near HII regions show higher intensities in all tracers. Galaxy centers, on the other hand, show higher overall intensities and enhanced CO-to-PAH ratios compared to galaxy disks. Individual galaxies show 0.19 dex scatter in the normalization of CO at fixed I_PAH and this normalization anti-correlates with specific star formation rate (SFR/M*) and correlates with stellar mass. We provide a prescription that accounts for these galaxy-to-galaxy variations and represents our best current empirical predictor to estimate CO(2-1) intensity from PAH emission, which allows one to take advantage of JWST's excellent sensitivity and resolution to trace cold gas., Comment: 21 pages, 4 figures, 3 tables. Submitted to ApJ

In this work, we aim to analyze and optimize the EnCLAP framework, a state-of-the-art model in automated audio captioning. We investigate the impact of modifying the acoustic encoder components, explore pretraining with different dataset scales, and study the effectiveness of a reranking scheme. Through extensive experimentation and quantitative analysis of generated captions, we develop EnCLAP++, an enhanced version that significantly surpasses the original., Comment: Accepted to DCASE2024 Workshop

In this technical report, we describe our submission to DCASE2024 Challenge Task6 (Automated Audio Captioning) and Task8 (Language-based Audio Retrieval). We develop our approach building upon the EnCLAP audio captioning framework and optimizing it for Task6 of the challenge. Notably, we outline the changes in the underlying components and the incorporation of the reranking process. Additionally, we submit a supplementary retriever model, a byproduct of our modified framework, to Task8. Our proposed systems achieve FENSE score of 0.542 on Task6 and mAP@10 score of 0.386 on Task8, significantly outperforming the baseline models., Comment: DCASE2024 Challenge Technical Report. Ranked 2nd in Task 6 Automated Audio Captioning

Recently, accelerated algorithms using the anchoring mechanism for minimax optimization and fixed-point problems have been proposed, and matching complexity lower bounds establish their optimality. In this work, we present the surprising observation that the optimal acceleration mechanism in minimax optimization and fixed-point problems is not unique. Our new algorithms achieve exactly the same worst-case convergence rates as existing anchor-based methods while using materially different acceleration mechanisms. Specifically, these new algorithms are dual to the prior anchor-based accelerated methods in the sense of H-duality. This finding opens a new avenue of research on accelerated algorithms since we now have a family of methods that empirically exhibit varied characteristics while having the same optimal worst-case guarantee.

We propose a framework to learn semantics from raw audio signals using two types of representations, encoding contextual and phonetic information respectively. Specifically, we introduce a speech-to-unit processing pipeline that captures two types of representations with different time resolutions. For the language model, we adopt a dual-channel architecture to incorporate both types of representation. We also present new training objectives, masked context reconstruction and masked context prediction, that push models to learn semantics effectively. Experiments on the sSIMI metric of Zero Resource Speech Benchmark 2021 and Fluent Speech Command dataset show our framework learns semantics better than models trained with only one type of representation., Comment: Accepted to ICASSP 2024

We propose EnCLAP, a novel framework for automated audio captioning. EnCLAP employs two acoustic representation models, EnCodec and CLAP, along with a pretrained language model, BART. We also introduce a new training objective called masked codec modeling that improves acoustic awareness of the pretrained language model. Experimental results on AudioCaps and Clotho demonstrate that our model surpasses the performance of baseline models. Source code will be available at https://github.com/jaeyeonkim99/EnCLAP . An online demo is available at https://huggingface.co/spaces/enclap-team/enclap ., Comment: Accepted to ICASSP 2024

The exquisite angular resolution and sensitivity of JWST is opening a new window for our understanding of the Universe. In nearby galaxies, JWST observations are revolutionizing our understanding of the first phases of star formation and the dusty interstellar medium. Nineteen local galaxies spanning a range of properties and morphologies across the star-forming main sequence have been observed as part of the PHANGS-JWST Cycle 1 Treasury program at spatial scales of $\sim$5-50pc. Here, we describe pjpipe, an image processing pipeline developed for the PHANGS-JWST program that wraps around and extends the official JWST pipeline. We release this pipeline to the community as it contains a number of tools generally useful for JWST NIRCam and MIRI observations. Particularly for extended sources, pjpipe products provide significant improvements over mosaics from the MAST archive in terms of removing instrumental noise in NIRCam data, background flux matching, and calibration of relative and absolute astrometry. We show that slightly smoothing F2100W MIRI data to 0.9" (degrading the resolution by about 30 percent) reduces the noise by a factor of $\approx$3. We also present the first public release (DR1.1.0) of the pjpipe processed eight-band 2-21 $\mu$m imaging for all nineteen galaxies in the PHANGS-JWST Cycle 1 Treasury program. An additional 55 galaxies will soon follow from a new PHANGS-JWST Cycle 2 Treasury program., Comment: 49 pages (27 in Appendices), 54 Figures (39 in Appendices), 3 Tables. Accepted for publication in ApJS. Updated to match accepted version. Data available at https://archive.stsci.edu/hlsp/phangs/phangs-jwst

We present the Physics at High Angular resolution in Nearby GalaxieS (PHANGS)-AstroSat atlas, which contains ultraviolet imaging of 31 nearby star-forming galaxies captured by the Ultraviolet Imaging Telescope (UVIT) on the AstroSat satellite. The atlas provides a homogeneous data set of far- and near-ultraviolet maps of galaxies within a distance of 22 Mpc and a median angular resolution of 1.4 arcseconds (corresponding to a physical scale between 25 and 160 pc). After subtracting a uniform ultraviolet background and accounting for Milky Way extinction, we compare our estimated flux densities to GALEX observations, finding good agreement. We find candidate extended UV disks around the galaxies NGC 6744 and IC 5332. We present the first statistical measurements of the clumping of the UV emission and compare it to the clumping of molecular gas traced with ALMA. We find that bars and spiral arms exhibit the highest degree of clumping, and the molecular gas is even more clumped than the FUV emission in galaxies. We investigate the variation of the ratio of observed FUV to H$\alpha$ in different galactic environments and kpc-sized apertures. We report that $\sim 65 \%$ varation of the $\log_{10}$(FUV/H$\alpha$) can be described through a combination of dust attenuation with star formation history parameters. The PHANGS-AstroSat atlas enhances the multi-wavelength coverage of our sample, offering a detailed perspective on star formation. When integrated with PHANGS data sets from ALMA, VLT-MUSE, HST and JWST, it develops our comprehensive understanding of attenuation curves and dust attenuation in star-forming galaxies., Comment: 35 pages, 16 figures. The survey webpage is available at https://sites.google.com/view/phangs/home/data/astrosat and the data archive can be accessed at https://www.canfar.net/storage/vault/list/phangs/RELEASES/PHANGS-AstroSat/v1p0

While first-order optimization methods are usually designed to efficiently reduce the function value $f(x)$, there has been recent interest in methods efficiently reducing the magnitude of $\nabla f(x)$, and the findings show that the two types of methods exhibit a certain symmetry. In this work, we present mirror duality, a one-to-one correspondence between mirror-descent-type methods reducing function value and reducing gradient magnitude. Using mirror duality, we obtain the dual accelerated mirror descent (dual-AMD) method that efficiently reduces $\psi^*(\nabla f(x))$, where $\psi$ is a distance-generating function and $\psi^*$ quantifies the magnitude of $\nabla f(x)$. We then apply dual-AMD to efficiently reduce $\|\nabla f(\cdot) \|_q$ for $q\in [2,\infty)$ and to efficiently compute $\varepsilon$-approximate solutions of the optimal transport problem.

Neural radiance field is an emerging rendering method that generates high-quality multi-view consistent images from a neural scene representation and volume rendering. Although neural radiance field-based techniques are robust for scene reconstruction, their ability to add or remove objects remains limited. This paper proposes a new language-driven approach for object manipulation with neural radiance fields through dataset updates. Specifically, to insert a new foreground object represented by a set of multi-view images into a background radiance field, we use a text-to-image diffusion model to learn and generate combined images that fuse the object of interest into the given background across views. These combined images are then used for refining the background radiance field so that we can render view-consistent images containing both the object and the background. To ensure view consistency, we propose a dataset updates strategy that prioritizes radiance field training with camera views close to the already-trained views prior to propagating the training to remaining views. We show that under the same dataset updates strategy, we can easily adapt our method for object insertion using data from text-to-3D models as well as object removal. Experimental results show that our method generates photorealistic images of the edited scenes, and outperforms state-of-the-art methods in 3D reconstruction and neural radiance field blending., Comment: CVPR 2024

In convex optimization, first-order optimization methods efficiently minimizing function values have been a central subject study since Nesterov's seminal work of 1983. Recently, however, Kim and Fessler's OGM-G and Lee et al.'s FISTA-G have been presented as alternatives that efficiently minimize the gradient magnitude instead. In this paper, we present H-duality, which represents a surprising one-to-one correspondence between methods efficiently minimizing function values and methods efficiently minimizing gradient magnitude. In continuous-time formulations, H-duality corresponds to reversing the time dependence of the dissipation/friction term. To the best of our knowledge, H-duality is different from Lagrange/Fenchel duality and is distinct from any previously known duality or symmetry relations. Using H-duality, we obtain a clearer understanding of the symmetry between Nesterov's method and OGM-G, derive a new class of methods efficiently reducing gradient magnitudes of smooth convex functions, and find a new composite minimization method that is simpler and faster than FISTA-G.

Previous pitch-controllable text-to-speech (TTS) models rely on directly modeling fundamental frequency, leading to low variance in synthesized speech. To address this issue, we propose PITS, an end-to-end pitch-controllable TTS model that utilizes variational inference to model pitch. Based on VITS, PITS incorporates the Yingram encoder, the Yingram decoder, and adversarial training of pitch-shifted synthesis to achieve pitch-controllability. Experiments demonstrate that PITS generates high-quality speech that is indistinguishable from ground truth speech and has high pitch-controllability without quality degradation. Code, audio samples, and demo are available at https://github.com/anonymous-pits/pits., Comment: 6 pages, preprint

JWST observations of polycyclic aromatic hydrocarbon (PAH) emission provide some of the deepest and highest resolution views of the cold interstellar medium (ISM) in nearby galaxies. If PAHs are well mixed with the atomic and molecular gas and illuminated by the average diffuse interstellar radiation field, PAH emission may provide an approximately linear, high resolution, high sensitivity tracer of diffuse gas surface density. We present a pilot study that explores using PAH emission in this way based on MIRI observations of IC 5332, NGC 628, NGC 1365, and NGC 7496 from the PHANGS-JWST Treasury. Using scaling relationships calibrated in Leroy et al. (2022), scaled F1130W provides 10--40 pc resolution and 3$\sigma$ sensitivity of $\Sigma_{\rm gas} \sim 2$ M$_\odot$ pc$^{-2}$. We characterize the surface densities of structures seen at $< 7$ M$_\odot$ pc$^{-2}$ in our targets, where we expect the gas to be HI-dominated. We highlight the existence of filaments, inter-arm emission, and holes in the diffuse ISM at these low surface densities. Below $\sim 10$ M$_\odot$ pc$^{-2}$ for NGC 628, NGC 1365, and NGC 7496 the gas distribution shows a ``Swiss cheese''-like topology due to holes and bubbles pervading the relatively smooth distribution of diffuse ISM. Comparing to recent galaxy simulations, we observe similar topology for the low surface density gas, though with notable variations between simulations with different setups and resolution. Such a comparison of high resolution, low surface density gas with simulations is not possible with existing atomic and molecular gas maps, highlighting the unique power of JWST maps of PAH emission., Comment: 18 pages, 6 figures, accepted as part of a PHANGS-JWST Focus Issue to appear in ApJ

We compare mid-infrared (mid-IR), extinction-corrected H$\alpha$, and CO (2-1) emission at 70--160 pc resolution in the first four PHANGS-JWST targets. We report correlation strengths, intensity ratios, and power law fits relating emission in JWST's F770W, F1000W, F1130W, and F2100W bands to CO and H$\alpha$. At these scales, CO and H$\alpha$ each correlate strongly with mid-IR emission, and these correlations are each stronger than the one relating CO to H$\alpha$ emission. This reflects that mid-IR emission simultaneously acts as a dust column density tracer, leading to the good match with the molecular gas-tracing CO, and as a heating tracer, leading to the good match with the H$\alpha$. By combining mid-IR, CO, and H$\alpha$ at scales where the overall correlation between cold gas and star formation begins to break down, we are able to separate these two effects. We model the mid-IR above $I_\nu = 0.5$~MJy sr$^{-1}$ at F770W, a cut designed to select regions where the molecular gas dominates the interstellar medium (ISM) mass. This bright emission can be described to first order by a model that combines a CO-tracing component and an H$\alpha$-tracing component. The best-fitting models imply that $\sim 50\%$ of the mid-IR flux arises from molecular gas heated by the diffuse interstellar radiation field, with the remaining $\sim 50\%$ associated with bright, dusty star forming regions. We discuss differences between the F770W, F1000W, F1130W bands and the continuum dominated F2100W band and suggest next steps for using the mid-IR as an ISM tracer., Comment: 49 pages, 17 figures, Section 8 provides a detailed summary, accepted for publication in ApJ Letters, part of a PHANGS-JWST Focus Issue to appear in ApJ

We explore the relationship between mid-infrared (mid-IR) and CO rotational line emission from massive star-forming galaxies, which is one of the tightest scalings in the local universe. We assemble a large set of unresolved and moderately ($\sim 1$ kpc) spatially resolved measurements of CO (1-0) and CO (2-1) intensity, $I_{\rm CO}$, and mid-IR intensity, $I_{\rm MIR}$, at 8, 12, 22, and 24$\mu$m. The $I_{\rm CO}$ vs. $I_{\rm MIR}$ relationship is reasonably described by a power law with slopes $0.7{-}1.2$ and normalization $I_{\rm CO} \sim 1$ K km s$^{-1}$ at $I_{\rm MIR} \sim 1$ MJy sr$^{-1}$. Both the slopes and intercepts vary systematically with choice of line and band. The comparison between the relations measured for CO~(1-0) and CO (2-1) allow us to infer that $R_{21} \propto I_{\rm MIR}^{0.2}$, in good agreement with other work. The $8\mu$m and $12\mu$m bands, with strong PAH features, show steeper CO vs. mid-IR slopes than the $22\mu$m and $24\mu$m, consistent with PAH emission arising not just from CO-bright gas but also from atomic or CO-dark gas. The CO-to-mid-IR ratio correlates with global galaxy stellar mass ($M_\star$) and anti-correlates with SFR/$M_\star$. At $\sim 1$ kpc resolution, the first four PHANGS-JWST targets show CO to mid-IR relationships that are quantitatively similar to our larger literature sample, including showing the steep CO-to-mid-IR slopes for the JWST PAH-tracing bands, although we caution that these initial data have a small sample size and span a limited range of intensities., Comment: 29 pages, 13 figures, key quantitative results in Table 3, Accepted as part of a PHANGS-JWST Focus Issue to appear in ApJ

Large-scale bars can fuel galaxy centers with molecular gas, often leading to the development of dense ring-like structures where intense star formation occurs, forming a very different environment compared to galactic disks. We pair ~0.3" (30pc) resolution new JWST/MIRI imaging with archival ALMA CO(2-1) mapping of the central ~5kpc of the nearby barred spiral galaxy NGC1365, to investigate the physical mechanisms responsible for this extreme star formation. The molecular gas morphology is resolved into two well-known bright bar lanes that surround a smooth dynamically cold gas disk (R_gal ~ 475pc) reminiscent of non-star-forming disks in early type galaxies and likely fed by gas inflow triggered by stellar feedback in the lanes. The lanes host a large number of JWST-identified massive young star clusters. We find some evidence for temporal star formation evolution along the ring. The complex kinematics in the gas lanes reveal strong streaming motions and may be consistent with convergence of gas streamlines expected there. Indeed, the extreme line-widths are found to be the result of inter-`cloud' motion between gas peaks; ScousePy decomposition reveals multiple components with line widths of ~ 19km/s and surface densities of ~ 800M_sun/pc^2, similar to the properties observed throughout the rest of the central molecular gas structure. Tailored hydro-dynamical simulations exhibit many of the observed properties and imply that the observed structures are transient and highly time-variable. From our study of NGC1365, we conclude that it is predominantly the high gas inflow triggered by the bar that is setting the star formation in its CMZ., Comment: 24 pages, 8 figures, accepted for publications as part of PHANGS-JWST ApJL Focus Issue

The PHANGS collaboration has been building a reference dataset for the multi-scale, multi-phase study of star formation and the interstellar medium in nearby galaxies. With the successful launch and commissioning of JWST, we can now obtain high-resolution infrared imaging to probe the youngest stellar populations and dust emission on the scales of star clusters and molecular clouds ($\sim$5-50 pc). In Cycle 1, PHANGS is conducting an 8-band imaging survey from 2-21$\mu$m of 19 nearby spiral galaxies. CO(2-1) mapping, optical integral field spectroscopy, and UV-optical imaging for all 19 galaxies have been obtained through large programs with ALMA, VLT/MUSE, and Hubble. PHANGS-JWST enables a full inventory of star formation, accurate measurement of the mass and age of star clusters, identification of the youngest embedded stellar populations, and characterization of the physical state of small dust grains. When combined with Hubble catalogs of $\sim$10,000 star clusters, MUSE spectroscopic mapping of $\sim$20,000 HII regions, and $\sim$12,000 ALMA-identified molecular clouds, it becomes possible to measure the timescales and efficiencies of the earliest phases of star formation and feedback, build an empirical model of the dependence of small dust grain properties on local ISM conditions, and test our understanding of how dust-reprocessed starlight traces star formation activity, all across a diversity of galactic environments. Here we describe the PHANGS-JWST Treasury survey, present the remarkable imaging obtained in the first few months of science operations, and provide context for the initial results presented in the first series of PHANGS-JWST publications., Comment: Re-submitted after addressing minor comments from referee. To be published as part of PHANGS-JWST ApJL Focus Issue

We use PHANGS-JWST data to identify and classify 1271 compact 21 $\mu$m sources in four nearby galaxies using MIRI F2100W data. We identify sources using a dendrogram-based algorithm, and we measure the background-subtracted flux densities for JWST bands from 2 $\mu$m to 21 $\mu$m. Using the SED in JWST as well as HST bands, plus ALMA and MUSE/VLT observations, we classify the sources by eye. Then we use this classification to define regions in color-color space, and so establish a quantitative framework for classifying sources. We identify 1085 sources as belonging to the ISM of the target galaxies with the remainder being dusty stars or background galaxies. These 21 $\mu$m sources are strongly spatially associated with HII regions ($>92\%$ of sources), while 74$\%$ of sources are coincident with a stellar association defined in the HST data. Using SED fitting, we find that the stellar masses of the 21 $\mu$m sources span a range of 10$^{2}$ to 10$^{4}~M_\odot$ with mass-weighted ages down to 2 Myr. There is a tight correlation between attenuation-corrected H$\alpha$ and 21 $\mu$m luminosity for $L_{\nu,\mathrm{F2100W}}>10^{19}~\mathrm{W~Hz}^{-1}$. Young embedded source candidates selected at 21 $\mu$m are found below this threshold and have $M_\star < 10^{3}~M_\odot$., Comment: 15 pages, 5 figures, Accepted as part of a PHANGS-JWST Focus Issue to appear in ApJ

We combine JWST observations with ALMA CO and VLT-MUSE H$\alpha$ data to examine off-spiral arm star formation in the face-on, grand-design spiral galaxy NGC 628. We focus on the northern spiral arm, around a galactocentric radius of 3-4 kpc, and study two spurs. These form an interesting contrast, as one is CO-rich and one CO-poor, and they have a maximum azimuthal offset in MIRI 21$\mu$m and MUSE H$\alpha$ of around 40$^\circ$ (CO-rich) and 55$^\circ$ (CO-poor) from the spiral arm. The star formation rate is higher in the regions of the spurs near to spiral arms, but the star formation efficiency appears relatively constant. Given the spiral pattern speed and rotation curve of this galaxy and assuming material exiting the arms undergoes purely circular motion, these offsets would be reached in 100-150 Myr, significantly longer than the 21$\mu$m and H$\alpha$ star formation timescales (both <10 Myr). The invariance of the star formation efficiency in the spurs versus the spiral arms indicates massive star formation is not only triggered in spiral arms, and cannot simply occur in the arms and then drift away from the wave pattern. These early JWST results show that in-situ star formation likely occurs in the spurs, and that the observed young stars are not simply the `leftovers' of stellar birth in the spiral arms. The excellent physical resolution and sensitivity that JWST can attain in nearby galaxies will well resolve individual star-forming regions and help us to better understand the earliest phases of star formation., Comment: 11 pages, 5 Figures, accepted as part of a PHANGS-JWST Focus Issue to appear in ApJ. Updated to include missing author and published paper references

The earliest stages of star formation, when young stars are still deeply embedded in their natal clouds, represent a critical phase in the matter cycle between gas clouds and young stellar regions. Until now, the high-resolution infrared observations required for characterizing this heavily obscured phase (during which massive stars have formed, but optical emission is not detected) could only be obtained for a handful of the most nearby galaxies. One of the main hurdles has been the limited angular resolution of the Spitzer Space Telescope. With the revolutionary capabilities of the JWST, it is now possible to investigate the matter cycle during the earliest phases of star formation as a function of the galactic environment. In this Letter, we demonstrate this by measuring the duration of the embedded phase of star formation and the implied time over which molecular clouds remain inert in the galaxy NGC628 at a distance of 9.8Mpc, demonstrating that the cosmic volume where this measurement can be made has increased by a factor of $>100$ compared to Spitzer. We show that young massive stars remain embedded for $5.1_{-1.4}^{+2.7}$Myr ($2.3_{-1.4}^{+2.7}$Myr of which being heavily obscured), representing $\sim20\%$ of the total cloud lifetime. These values are in broad agreement with previous measurements in five nearby ($D < 3.5$Mpc) galaxies and constitute a proof of concept for the systematic characterization of the early phase of star formation across the nearby galaxy population with the PHANGS-JWST survey., Comment: 15 pages, 4 figures, 2 tables; accepted for publication in ApJL as part of a PHANGS-JWST First Results Focus issue

The earliest stages of star formation occur enshrouded in dust and are not observable in the optical. Here we leverage the extraordinary new high-resolution infrared imaging from JWST to begin the study of dust-embedded star clusters in nearby galaxies throughout the local volume. We present a technique for identifying dust-embedded clusters in NGC 7496 (18.7 Mpc), the first galaxy to be observed by the PHANGS-JWST Cycle 1 Treasury Survey. We select sources that have strong 3.3$\mu$m PAH emission based on a $\rm F300M-F335M$ color excess, and identify 67 candidate embedded clusters. Only eight of these are found in the PHANGS-HST optically-selected cluster catalog and all are young (six have SED-fit ages of $\sim1$ Myr). We find that this sample of embedded cluster candidates may significantly increase the census of young clusters in NGC 7496 from the PHANGS-HST catalog -- the number of clusters younger than $\sim$2 Myr could be increased by a factor of two. Candidates are preferentially located in dust lanes, and are coincident with peaks in PHANGS-ALMA CO (2-1) maps. We take a first look at concentration indices, luminosity functions, SEDs spanning from 2700A to 21$\mu$m, and stellar masses (estimated to be between $\sim10^4-10^5 M_{\odot}$). The methods tested here provide a basis for future work to derive accurate constraints on the physical properties of embedded clusters, characterize the completeness of cluster samples, and expand analysis to all 19 galaxies in the PHANGS-JWST sample, which will enable basic unsolved problems in star formation and cluster evolution to be addressed., Comment: 12 pages, 6 figures; accepted for publication in ApJL as part of PHANGS-JWST First Results Special Issue

In this paper, we propose a new method for mapping a 3D point cloud to the latent space of a 3D generative adversarial network. Our generative model for 3D point clouds is based on SP-GAN, a state-of-the-art sphere-guided 3D point cloud generator. We derive an efficient way to encode an input 3D point cloud to the latent space of the SP-GAN. Our point cloud encoder can resolve the point ordering issue during inversion, and thus can determine the correspondences between points in the generated 3D point cloud and those in the canonical sphere used by the generator. We show that our method outperforms previous GAN inversion methods for 3D point clouds, achieving state-of-the-art results both quantitatively and qualitatively. Our code is available at https://github.com/hkust-vgd/point_inverter., Comment: WACV 2023 paper. 8 pages of main content, 2 pages of references, 7 pages of supplementary material

The time-scales associated with various stages of the star formation process represent major unknowns in our understanding of galactic evolution, as well as of star and planet formation. This is the second paper in a series aiming to establish a multi-tracer time-line of star formation in the Large Magellanic Cloud (LMC), focusing on the lifecycle of molecular clouds. We use a statistical method to determine a molecular cloud lifetime in the LMC of $t_{\text{CO}}=11.8^{+2.7}_{-2.2}$ Myr. This short time-scale is similar to the cloud dynamical time, and suggests that molecular clouds in the LMC are largely decoupled from the effects of galactic dynamics and have lifetimes set by internal processes. This provides a clear contrast to atomic clouds in the LMC, of which the lifetimes are correlated with galactic dynamical time-scales. We additionally derive the time-scale for which molecular clouds and HII regions co-exist as $t_{\text{fb}}=1.2^{+0.3}_{-0.2}$ Myr, implying an average feedback front expansion velocity of 12 km s$^{-1}$, consistent with expansion velocities of HII regions in the LMC observed directly using optical spectroscopy. Taken together, these results imply that the molecular cloud lifecycle in the LMC proceeds rapidly and is regulated by internal dynamics and stellar feedback. We conclude by discussing our measurements in the context of previous work in the literature, which reported considerably longer lifetimes for molecular clouds in the LMC, and find that these previous findings resulted from a subjective choice in timeline calibration that is avoided by our statistical methodology., Comment: 18 pages (including appendices), 10 figures, 2 tables; MNRAS in press (accepted August 15, 2022). Figures 3, 4, 5, and 7 show the main results of the paper

The processes of star formation and feedback, regulating the cycle of matter between gas and stars on the scales of giant molecular clouds (GMCs; $\sim$100pc), play a major role in governing galaxy evolution. Measuring the time-scales of GMC evolution is important to identify and characterise the specific physical mechanisms that drive this transition. By applying a robust statistical method to high-resolution CO and narrow-band H$\alpha$ imaging from the PHANGS survey, we systematically measure the evolutionary timeline from molecular clouds to exposed young stellar regions on GMC scales, across the discs of an unprecedented sample of 54 star-forming main-sequence galaxies (excluding their unresolved centres). We find that clouds live for about $1{-}3$ GMC turbulence crossing times ($5{-}30$Myr) and are efficiently dispersed by stellar feedback within $1{-}5$Myr once the star-forming region becomes partially exposed, resulting in integrated star formation efficiencies of $1{-}8$%. These ranges reflect physical galaxy-to-galaxy variation. In order to evaluate whether galactic environment influences GMC evolution, we correlate our measurements with average properties of the GMCs and their local galactic environment. We find several strong correlations that can be physically understood, revealing a quantitative link between galactic-scale environmental properties and the small-scale GMC evolution. Notably, the measured CO-visible cloud lifetimes become shorter with decreasing galaxy mass, mostly due to the increasing presence of CO-dark molecular gas in such environment. Our results represent a first step towards a comprehensive picture of cloud assembly and dispersal, which requires further extension and refinement with tracers of the atomic gas, dust, and deeply-embedded stars., Comment: 24 pages, 8 figures, 5 tables; accepted for publication in MNRAS (August 16, 2022)

High-resolution daytime satellite imagery has become a promising source to study economic activities. These images display detailed terrain over large areas and allow zooming into smaller neighborhoods. Existing methods, however, have utilized images only in a single-level geographical unit. This research presents a deep learning model to predict economic indicators via aggregating traits observed from multiple levels of geographical units. The model first measures hyperlocal economy over small communities via ordinal regression. The next step extracts district-level features by summarizing interconnection among hyperlocal economies. In the final step, the model estimates economic indicators of districts via aggregating the hyperlocal and district information. Our new multi-level learning model substantially outperforms strong baselines in predicting key indicators such as population, purchasing power, and energy consumption. The model is also robust against data shortage; the trained features from one country can generalize to other countries when evaluated with data gathered from Malaysia, the Philippines, Thailand, and Vietnam. We discuss the multi-level model's implications for measuring inequality, which is the essential first step in policy and social science research on inequality and poverty., Comment: Accepted at AAAI2022

We use an unprecedented sample of about 23,000 HII regions detected at an average physical resolution of 67pc in the PHANGS-MUSE sample to study the extragalactic HII region Ha luminosity function (LF). Our observations probe the star-forming disk of 19 nearby spiral galaxies with low inclination and located close to the star formation main sequence at z=0. The mean LF slope $\alpha$ in our sample is =1.73 with a $\sigma$ of 0.15. We find that $\alpha$ decreases with the galaxy's star formation rate surface density and argue that this is driven by an enhanced clustering of young stars at high gas surface densities. Looking at the HII regions within single galaxies we find that no significant variations occur between the LF of the inner and outer part of the star-forming disk, whereas the LF in the spiral arm areas is shallower than in the inter-arm areas for six out of the 13 galaxies with clearly visible spiral arms. We attribute these variations to the spiral arms increasing the molecular clouds arm--inter-arm mass contrast and find suggestive evidence that they are more evident for galaxies with stronger spiral arms. Furthermore, we find systematic variations in $\alpha$ between samples of HII regions with high and low ionization parameter q and argue that they are driven by the aging of HII regions., Comment: Accepted in A&A. Main text: 19 pages, 11 Figures, 4 Tables. Catalog including 23301 HII regions will be available in electronic form at the CDS upon official publication

We performed low-resolution spectroscopy for the red giant branch stars in an intriguing globular cluster (GC) NGC 2808, which hosts subpopulations with extreme helium and light-element abundances. In order to trace N, C, and Ca abundance differences among subpopulations, we measured CN, CH, and Ca II H&K spectral indices, respectively. We identified four subpopulations (G1, G2, G3, and G4) from CN and CH strength, with CN-weak/CH-strong G1, CN-intermediate/CH-strong G2, CN-strong/CH-intermediate G3, and CN-strong/CH-weak G4. Compared to [Na/O] from high-resolution spectroscopy, we show that CN index can more clearly separate G1 and G2. Since CN traces N abundance in a GC, it implies that G1 and G2 would show a larger difference in [N/Fe] compared to [Na/Fe], as predicted by chemical evolution models. Later generation stars G3 and G4, however, are better separated with high-resolution spectroscopy. We also found that G4 shows a stronger Ca II H&K line strength compared to that of G1, but we suspect this to be a result of unusually strong He enhancement and/or Mg depletion in G4 of this GC. This work illustrates that combining low- and high-resolution spectroscopic studies can improve the separation of subpopulations in GCs., Comment: Accepted for publication in AJ

We present PHANGS-ALMA, the first survey to map CO J=2-1 line emission at ~1" ~ 100pc spatial resolution from a representative sample of 90 nearby (d<~20 Mpc) galaxies that lie on or near the z=0 "main sequence" of star-forming galaxies. CO line emission traces the bulk distribution of molecular gas, which is the cold, star-forming phase of the interstellar medium. At the resolution achieved by PHANGS-ALMA, each beam reaches the size of a typical individual giant molecular cloud (GMC), so that these data can be used to measure the demographics, life-cycle, and physical state of molecular clouds across the population of galaxies where the majority of stars form at z=0. This paper describes the scientific motivation and background for the survey, sample selection, global properties of the targets, ALMA observations, and characteristics of the delivered ALMA data and derived data products. As the ALMA sample serves as the parent sample for parallel surveys with VLT/MUSE, HST, AstroSat, VLA, and other facilities, we include a detailed discussion of the sample selection. We detail the estimation of galaxy mass, size, star formation rate, CO luminosity, and other properties, compare estimates using different systems and provide best-estimate integrated measurements for each target. We also report the design and execution of the ALMA observations, which combine a Cycle~5 Large Program, a series of smaller programs, and archival observations. Finally, we present the first 1" resolution atlas of CO emission from nearby galaxies and describe the properties and contents of the first PHANGS-ALMA public data release., Comment: 76 pages, 33 figures. Accepted for publication in the Astrophysical Journal Supplement series. Full resolution version and the image atlas to appear as a figure set in the published version can be found https://sites.google.com/view/phangs/publications . Data release coming soon to the ALMA archive and CADC temporarily available at http://phangs.org/data

Feedback from massive stars plays a key role in molecular cloud evolution. After the onset of star formation, the young stellar population is exposed by photoionization, winds, supernovae, and radiation pressure from massive stars. Recent observations of nearby galaxies have provided the evolutionary timeline between molecular clouds and exposed young stars, but the duration of the embedded phase of massive star formation is still ill-constrained. We measure how long massive stellar populations remain embedded within their natal cloud, by applying a statistical method to six nearby galaxies at 20-100 pc resolution, using CO, Spitzer 24$\rm\,\mu m$, and H$\alpha$ emission as tracers of molecular clouds, embedded star formation, and exposed star formation, respectively. We find that the embedded phase (with CO and 24$\rm\,\mu m$ emission) lasts for $2{-}7$ Myr and constitutes $17{-}47\%$ of the cloud lifetime. During approximately the first half of this phase, the region is invisible in H$\alpha$, making it heavily obscured. For the second half of this phase, the region also emits in H$\alpha$ and is partially exposed. Once the cloud has been dispersed by feedback, 24$\rm\,\mu m$ emission no longer traces ongoing star formation, but remains detectable for another $2{-}9$ Myr through the emission from ambient CO-dark gas, tracing star formation that recently ended. The short duration of massive star formation suggests that pre-supernova feedback (photoionization and winds) is important in disrupting molecular clouds. The measured timescales do not show significant correlations with environmental properties (e.g. metallicity). Future JWST observations will enable these measurements routinely across the nearby galaxy population., Comment: 23 pages, 9 figures, 4 tables; accepted for publication in MNRAS (March 19, 2021)

With recent developments of convolutional neural networks, deep learning for 3D point clouds has shown significant progress in various 3D scene understanding tasks, e.g., object recognition, semantic segmentation. In a safety-critical environment, it is however not well understood how such deep learning models are vulnerable to adversarial examples. In this work, we explore adversarial attacks for point cloud-based neural networks. We propose a unified formulation for adversarial point cloud generation that can generalise two different attack strategies. Our method generates adversarial examples by attacking the classification ability of point cloud-based networks while considering the perceptibility of the examples and ensuring the minimal level of point manipulations. Experimental results show that our method achieves the state-of-the-art performance with higher than 89% and 90% of attack success rate on synthetic and real-world data respectively, while manipulating only about 4% of the total points., Comment: ICCV 2021 camera-ready paper (8 pages)