Search

Large pre-trained models have achieved outstanding results in sequence modeling. The Transformer block and its attention mechanism have been the main drivers of the success of these models. Recently, alternative architectures, such as Selective Structured State Space Models (SSMs), have been proposed to address the inefficiencies of Transformers. This paper explores the compression of SSM-based models, particularly Mamba and its hybrids. We study the sensitivity of these models to the removal of selected components at different granularities to reduce the model size and computational overhead, thus improving their efficiency while maintaining accuracy. The proposed solutions, collectively referred to as Mamba-Shedder, achieve a speedup of up to 1.4x during inference, demonstrating that model efficiency can be improved by eliminating several redundancies with minimal impact on the overall model performance. The code is available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning., Comment: NAACL-25 - Main track

The rapid expansion of Large Language Models (LLMs) has posed significant challenges regarding the computational resources required for fine-tuning and deployment. Recent advancements in low-rank adapters have demonstrated their efficacy in parameter-efficient fine-tuning (PEFT) of these models. This retrospective paper comprehensively discusses innovative approaches that synergize low-rank representations with Neural Architecture Search (NAS) techniques, particularly weight-sharing super-networks. Robust solutions for compressing and fine-tuning large pre-trained models are developed by integrating these methodologies. Our analysis highlights the potential of these combined strategies to democratize the use of LLMs, making them more accessible for deployment in resource-constrained environments. The resulting models exhibit reduced memory footprints and faster inference times, paving the way for more practical and scalable applications of LLMs. Models and code are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning., Comment: AAAI-25 Workshop on Connecting Low-rank Representations in AI

Recently, state-of-the-art approaches for pruning large pre-trained models (LPMs) have demonstrated that the training-free removal of non-critical residual blocks in Transformers is viable for reducing model size, achieving results that outperform previous training-free pruning approaches. Motivated by these findings, we extend BlockPruner (Zhong et al., 2024) and propose MultiPruner, a pruning approach that surpasses recent training-free pruning methods by adopting a multidimensional, iterative, fine-grained pruning strategy. In MultiPruner, multidimensional pruning reinstates the structural balance in block-pruned models by sequentially compressing along three dimensions: i) residual blocks, ii) channels of multilayer perceptrons (MLP), and iii) attention heads. This solution enhances zero-shot accuracy on downstream tasks compared to other techniques while improving model compression ratios, producing compressed models with fewer computing and memory requirements. Extensive experiments demonstrate the advantages of the proposed method across various large pre-trained models. The code and pruning configurations are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.

Context. The gravitational lens system SDSS J1004+4112 was the first known example of a quasar lensed by a galaxy cluster. The interest in this system has been renewed following the publication of r-band light curves spanning 14.5 years and the determination of the time delays between the four brightest quasar images. Aims. We constrained the quasar accretion disk size and the fraction of the lens mass in stars using the signature of microlensing in the quasar image light curves. Methods. We built the six possible histograms of microlensing magnitude differences between the four quasar images and compared them with simulated model histograms, using a $\chi^2$ test to infer the model parameters. Results. We infer a quasar disk half-light radius of $R_{1/2}=(0.70\pm0.04)\, R_E=(6.4\pm0.4) \sqrt{M/0.3M_{\odot}}$ light-days at 2407\r{A} in the rest frame and stellar mass fractions at the quasar image positions of $\alpha_A>0.059$, $\alpha_B=0.056^{+0.021}_{-0.027}$, $\alpha_C=0.030^{+0.031}_{-0.021}$, and $\alpha_D=0.072^{+0.034}_{-0.016}$. Conclusions. The inferred disk size is broadly compatible with most previous estimates, and the stellar mass fractions are within the expected ranges for galaxy clusters. In the region where image C lies, the stellar mass fraction is compatible with a stellar contribution from the brightest cluster galaxy, galaxy cluster members, and intracluster light, but the values at images B, D, and especially A are slightly larger, possibly suggesting the presence of extra stellar components., Comment: 7 pages, 6 figures, 2 tables, A&A in press

We show that the spin of dark matter induces a gravitational analog of the electromagnetic Faraday effect, where the polarization of gravitational waves undergoes a rotation as they propagate through a dark matter halo with a non-vanishing axial spin tensor (hypermomentum). An expression for the gravitational rotation angle is provided, which is analogous to the Faraday rotation in optics, and evaluate its significance in astrophysical settings. Although the effect is expected to be small under current observational constraints, we discuss its potential importance in the early universe.

We study the properties of a cryogenic buffer gas source that uses a low temperature two-stage buffer gas cell to produce very slow beams of ytterbium monofluoride molecules. The molecules are produced by laser ablation inside the cell and extracted into a beam by a flow of cold helium. We measure the flux and velocity distribution of the beam as a function of ablation energy, helium flow rate, cell temperature, and the size of the gap between the first and second stages of the cell. We also compare the velocity distributions from one-stage and two-stage cells. The one-stage cell emits a beam with a speed of about 82 m s$^{-1}$ and a translational temperature of 0.63 K. The slowest beams are obtained using the two-stage cell at the lowest achievable cell temperature of 1.8 K. This beam has a peak velocity of 56 m s$^{-1}$ and a flux of $9 \times 10^9$ ground state molecules per steradian per pulse, with a substantial fraction at speeds below 40 m s$^{-1}$. These slow molecules can be decelerated further by radiation pressure slowing and then captured in a magneto-optical trap., Comment: 11 pages, 8 figures; Extra figure added and some improvements to the text throughout

This study provides an analysis of atmospheric neutrino oscillations at the ESSnuSB far detector facility. The prospects of the two cylindrical Water Cherenkov detectors with a total fiducial mass of 540 kt are investigated over 10 years of data taking in the standard three-flavor oscillation scenario. We present the confidence intervals for the determination of mass ordering, $\theta_{23}$ octant as well as for the precisions on $\sin^2\theta_{23}$ and $|\Delta m_{31}^2|$. It is shown that mass ordering can be resolved by $3\sigma$ CL ($5\sigma$ CL) after 4 years (10 years) regardless of the true neutrino mass ordering. Correspondingly, the wrong $\theta_{23}$ octant could be excluded by $3\sigma$ CL after 4 years (8 years) in the case where the true neutrino mass ordering is normal ordering (inverted ordering). The results presented in this work are complementary to the accelerator neutrino program in the ESSnuSB project., Comment: 14 pages, 7 figures and 2 tables, accepted for publication in the Journal of High Energy Physics

We introduce a novel method to derive rotation curves with light-day spatial resolution of the inner regions of lensed quasars. We aim to probe the kinematics of the inner part of the broad-line region (BLR) by resolving the microlensing response - a proxy for the size of the emitting region - in the wings of the broad emission lines (BELs). Specifically, we assess the strength of the microlensing effects in the wings of the high-ionization lines Si IV and C IV across various velocity bins in five gravitationally lensed quasars: SDSS J1001+5027, SDSS J1004+4112, HE 1104$-$1805, SDSS J1206+4332, and SDSS J1339+1310. Using Bayesian methods to estimate the dimensions of the corresponding emission regions and adopting a Keplerian model as our baseline, we examine the consistency of the hypothesis of disk-like rotation. Our results reveal a monotonic, smooth increase in microlensing magnification with velocity. The deduced velocity-size relationships inferred for the various quasars and emission lines closely conform to the Keplerian model of an inclined disk. This study provides the first direct evidence of Keplerian rotation in the innermost region of quasars across a range of radial distances spanning from $\sim$5 to 20 light-days.

We aim to unveil the structure of the continuum and broad-emission line (BEL) emitting regions in the gravitationally lensed quasar SDSS J1339+1310 by examining the distinct signatures of microlensing present in this system. Our study involves a comprehensive analysis of ten years (2009-2019) of photometric monitoring data and seven spectroscopic observations acquired between 2007 and 2017. This work focuses on the pronounced deformations in the BEL profiles between images A and B, alongside the chromatic changes in their adjacent continua and the striking microlensing variability observed in the $r$-band light curves. We employed a statistical model to quantify the distribution and impact of microlensing magnifications and utilized a Bayesian approach to estimate the dimensions of various emission regions within the quasar. The analysis of the $r$-band light curves reveals substantial microlensing variability in the rest-frame UV continuum, suggesting that image B is amplified relative to image A by a factor of up to six. This finding is corroborated by pronounced microlensing-induced distortions in all studied BEL profiles (Ly$\alpha$, Si IV, C IV, C III], and Mg II), especially a prominent magnification of image B's red wing. We estimated the average dimensions of the BLR to be notably smaller than usual: the region emitting the blue wings measures $R_{1/2} = 11.5 \pm 1.7$ light-days, while the red wings originate from a more compact area of $R_{1/2} = 2.9\pm0.6$ light-days. From the photometric monitoring data, we inferred that the region emitting the $r$-band is $R_{1/2} = 2.2\pm0.3$ light-days across. Furthermore, by assessing the gravitational redshift of the UV Fe III blend and combining it with the blend's microlensing-based size estimate, we calculated the central SMBH's mass to be $M_{BH} \sim2 \times 10^8 M_\odot$.

We propose and discuss a new experimental approach to measure the centroid shift induced by gravitational microlensing in the images of lensed quasars (astrometric microlensing). Our strategy is based on taking the photocenter of a region in the quasar large enough as to be insensitive to microlensing as reference to measure the centroid displacement of the continuum. In this way, single-epoch measurements of astrometric microlensing can be performed. Using numerical simulations, we show that, indeed, the centroid shift monotonically decreases as the size of the emitting region increases, and only for relatively large regions, like the broad line region (BLR), does the centroid shift become negligible. This opens interesting possibilities to study the stratification of the different emitters in the accretion disk and the BLR. We estimate the amplitude of the centroid shifts for 79 gravitationally lensed images and study more thoroughly the special cases Q2237+030 A, RXJ1131-1231 A, PG1115+080 A2 and SDSS J1004+4112 A. We propose to use spectro-astrometry to simultaneously obtain the photocenters of the continuum and of different emission line regions since, with the precision of forthcoming instruments, astrometric microlensing by $\sim 1 M_\odot$ mass microlenses may be detected in many quasar lensed images. When we consider more massive micro/millilenses, $M\gtrsim 10 M_\odot$, often proposed as the constituents of dark matter, the BLR becomes sensitive to microlensing and can no longer be used as a positional reference to measure centroid shifts. Differential microlensing between the images of a lensed quasar along several epochs should be used instead., Comment: 14 pages, 6 figures, 1 table. ApJ in press

Neutrino oscillation experiments provide a unique window in exploring several new physics scenarios beyond the standard three flavour. One such scenario is quantum decoherence in neutrino oscillation which tends to destroy the interference pattern of neutrinos reaching the far detector from the source. In this work, we study the decoherence in neutrino oscillation in the context of the ESSnuSB experiment. We consider the energy-independent decoherence parameter and derive the analytical expressions for P$_{\mu e}$ and P$_{\mu \mu}$ probabilities in vacuum. We have computed the capability of ESSnuSB to put bounds on the decoherence parameters namely, $\Gamma_{21}$ and $\Gamma_{32}$ and found that the constraints on $\Gamma_{21}$ are competitive compared to the DUNE bounds and better than the most stringent LBL ones from MINOS/MINOS+. We have also investigated the impact of decoherence on the ESSnuSB measurement of the Dirac CP phase $\delta_{\rm CP}$ and concluded that it remains robust in the presence of new physics., Comment: 30 pages, 9 figures, 2 tables, Version accepted for publication in JHEP

Recently, several approaches successfully demonstrated that weight-sharing Neural Architecture Search (NAS) can effectively explore a search space of elastic low-rank adapters (LoRA), allowing the parameter-efficient fine-tuning (PEFT) and compression of large language models. In this paper, we introduce a novel approach called Shears, demonstrating how the integration of cost-effective sparsity and a proposed Neural Low-rank adapter Search (NLS) algorithm can further improve the efficiency of PEFT approaches. Results demonstrate the benefits of Shears compared to other methods, reaching high sparsity levels while improving or with little drop in accuracy, utilizing a single GPU for a pair of hours., Comment: 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics (Industry Track)

In the rapidly evolving field of artificial intelligence, multimodal models, e.g., integrating vision and language into visual-language models (VLMs), have become pivotal for many applications, ranging from image captioning to multimodal search engines. Among these models, the Contrastive Language-Image Pre-training (CLIP) model has demonstrated remarkable performance in understanding and generating nuanced relationships between text and images. However, the conventional training of such models often requires centralized aggregation of vast datasets, posing significant privacy and data governance challenges. To address these concerns, this paper proposes a novel approach that leverages Federated Learning and parameter-efficient adapters, i.e., Low-Rank Adaptation (LoRA), to train VLMs. This methodology preserves data privacy by training models across decentralized data sources and ensures model adaptability and efficiency through LoRA's parameter-efficient fine-tuning. Our approach accelerates training time by up to 34.72 times and requires 2.47 times less memory usage than full fine-tuning., Comment: 10 pages, 11 figures

We employ a lattice Boltzmann method to compute the acoustic radiation force produced by standing waves on a compressible object. Instead of simulating the fluid mechanics equations directly, the proposed method uses a lattice Boltzmann model that reproduces the wave equation, together with a kernel interpolation scheme, to compute the first order perturbations of the pressure and velocity fields on the object's surface and, from them, the acoustic radiation force. The procedure reproduces with excellent accuracy the theoretical expressions by Gor'kov and Wei for the sphere and the disk, respectively, even with a modest number of lattice Boltzmann cells. The proposed method shows to be a promising tool for simulating phenomena where the acoustic radiation force plays a relevant role, like acoustic tweezers or the acoustic manipulation of microswimmers, with applications in medicine and engineering.

Lithium ferrites are well known materials due to their numerous technological applications especially in microwave devices. Mn-doped lithium ferrite nanoparticles were prepared by sol-gel technique by means of Pechini method, and then annealed at different temperatures in 250 to 1000 {\deg}C range. XRD confirms spinel formation with particle size in the 15 to 200 nm range, with increased size with annealing temperature. Microwave magnetoabsorption data of annealed lithium ferrite nanoparticles, obtained with a broadband system based on a network analyzer operating up to 8.5 GHz are presented. At fields up to 200 mT we can observe a broad absorption peak that shifts to higher frequencies with magnetic field according to ferromagnetic resonance theory. The amplitude of absorption, up to 85%, together with the frequency width of about 4.5 GHz makes this material suitable as wave absorber. Samples annealed at higher temperatures show a behaviour similar to polycrystalline samples, thus suggesting their multidomain character

Single phase multiferroics in which ordered magnetic and ferroelectricity coexist, are of great interest for new multifunctional devices, and among them Y-type hexaferrites are good candidates. Transverse susceptibility measurements, which have been proved to be a versatile tool to study singular properties of bulk and nanoparticle magnetic systems, have been carried out with a broadband system on polycrystalline Y type hexaferrites with composition Ba$_{0.5}$Sr$_{1.5}$Co$_2$Fe$_{12}$O${22}$, optimal to exhibit multiferroic properties. In the temperature range 80-350 K transverse susceptibility measurements with DC fields up to $\pm$5000 Oe reveal different behaviour depending on the sintering temperature. The thermal evolution of the anisotropy field peak exhibit four regions with different slopes: positive in 80-130 K, negative in 130-200 K, constant in 200-280 K and negative in 280-350 K, which can be considered a signature of spin transitions in this compound., Comment: arXiv admin note: substantial text overlap with arXiv:2401.16146

Neutron transmission experiments can offer a new type of highly sensitive search for time-reversal invariance violating (TRIV) effects in nucleon-nucleon interactions via the same enhancement mechanism observed for large parity violating (PV) effects in neutron-induced compound nuclear processes. In these compound processes, the TRIV cross-section is given as the product of the PV cross-section, a spin-factor $\kappa$, and a ratio of TRIV and PV matrix elements. We determined $\kappa$ to be $0.59\pm0.05$ for $^{139}$La+$n$ using both $(n, \gamma)$ spectroscopy and ($\vec{n}+^{139}\vec{\rm La}$) transmission. This result quantifies for the first time the high sensitivity of the $^{139}$La 0.75~eV $p$-wave resonance in a future search for P-odd/T-odd interactions in ($\vec{n}+^{139}\vec{\rm La}$) forward transmission.

In this paper, we delve into the intricacies of boundary stabilization for the Linearized KP-II Equation within the constraints of a bounded domain. Our primary aim is to engineer a parameterized set of feedback laws that not only lead to the existence of solutions originating from small enough initial data but also facilitate their uniform exponential decay within the energy space. As far as we know, the present work is the first result on boundary feedback stabilization related to the Kadomtsev-Petviashvili II equation posed in a rectangle $\Omega=(0,L)\times (0,L)$, $L>0$., Comment: 14 pages, preprint

We present the characterisation of an inner mini-Neptune in a 9.2292005$\pm$0.0000063 day orbit and an outer mono-transiting sub-Saturn planet in a 95.50$^{+0.36}_{-0.25}$ day orbit around the moderately active, bright (mv=8.9 mag) K5V star TOI-2134. Based on our analysis of five sectors of TESS data, we determine the radii of TOI-2134b and c to be 2.69$\pm$0.16 R$_{e}$ for the inner planet and 7.27$\pm$0.42 R$_{e}$ for the outer one. We acquired 111 radial-velocity spectra with HARPS-N and 108 radial-velocity spectra with SOPHIE. After careful periodogram analysis, we derive masses for both planets via Gaussian Process regression: 9.13$^{+0.78}_{-0.76}$ M$_{e}$ for TOI-2134b and 41.86$^{+7.69}_{-7.83}$ M$_{e}$ for TOI-2134c. We analysed the photometric and radial-velocity data first separately, then jointly. The inner planet is a mini-Neptune with density consistent with either a water-world or a rocky core planet with a low-mass H/He envelope. The outer planet has a bulk density similar to Saturn's. The outer planet is derived to have a significant eccentricity of 0.67$^{+0.05}_{-0.06}$ from a combination of photometry and RVs. We compute the irradiation of TOI-2134c as 1.45$\pm$0.10 times the bolometric flux received by Earth, positioning it for part of its orbit in the habitable sone of its system. We recommend further RV observations to fully constrain the orbit of TOI-2134c. With an expected Rossiter-McLaughlin (RM) effect amplitude of 7.2$\pm$1.3 m/s, we recommend TOI-2134c for follow-up RM analysis to study the spin-orbit architecture of the system. We calculate the Transmission Spectroscopy Metric, and both planets are suitable for bright-mode NIRCam atmospheric characterisation., Comment: 24 pages, 7 tables, 14 figures

We aim to reveal the sizes of the continuum and broad emission line (BEL) emitting regions in the gravitationally lensed quasar SDSS J1004+4112 by analyzing the unique signatures of microlensing in this system. Through a comprehensive analysis of 20 spectroscopic observations acquired between 2003 and 2018, we studied the striking deformations of various BEL profiles and determined the sizes of their respective emitting regions. Our approach involves a detailed analysis of the magnitude differences in the BEL wings and their adjacent continua, and the implementation of a statistical model to quantify the distribution and impact of microlensing magnifications. To ensure a reliable baseline for no microlensing, we used the emission line cores as a reference. We then applied a Bayesian estimate to derive the size lower limits of the Ly$\alpha$, Si IV, C IV, C III], and Mg II emitting regions, as well as the sizes of the underlying continuum-emitting sources. We analyzed the outstanding microlensing-induced distortions in the line profiles of various BELs in the quasar image A, characterized by a prominent magnification of the blue part and a strong demagnification of the red part. From the statistics of microlensing magnifications and using Bayesian methods, we estimate the lower limit to the overall size of the regions emitting the BELs to be a few lt-days across, which is significantly smaller than in typically lensed quasars. The asymmetric deformations in the BELs indicate that the broad-line region is generally not spherically symmetric, and is likely confined to a plane and following the motions of the accretion disk. Additionally, the inferred continuum-emitting region sizes are larger than predictions based on standard thin-disk theory by a factor of $\sim$3.6 on average. The size-wavelength relation is consistent with that of a geometrically thin and optically thick accretion disk.

In this paper we study non-standard interactions mediated by a scalar field (SNSI) in the context of ESSnuSB experiment. In particular we study the capability of ESSnuSB to put bounds on the SNSI parameters and also study the impact of SNSI in the measurement of the leptonic CP phase $\delta_{\rm CP}$. Existence of SNSI modifies the neutrino mass matrix and this modification can be expressed in terms of three diagonal real parameters ($\eta_{ee}$, $\eta_{\mu\mu}$ and $\eta_{\tau\tau}$) and three off-diagonal complex parameters ($\eta_{e \mu}$, $\eta_{e\tau}$ and $\eta_{\mu\tau}$). Our study shows that the upper bounds on the parameters $\eta_{\mu\mu}$, $\eta_{\tau\tau}$ and $\eta_{\mu\tau}$ depend upon how $\Delta m^2_{31}$ is minimized in the theory. However, this is not the case when one tries to measure the impact of SNSI on $\delta_{\rm CP}$. Further, we show that the CP sensitivity of ESSnuSB can be completely lost for certain values of $\eta_{ee}$ and $\eta_{\mu\tau}$ for which the appearance channel probability becomes independent of $\delta_{\rm CP}$., Comment: 14 pages, 6 figures, 2 tables, Version accepted for publication in Phys. Rev. D

In this article we revisit the projectile motion assuming a retarding force proportional to the velocity, $\vec{F_r} = -mk\vec{V}$. We obtain an analytical expression for the set of maxima of the trajectories, in Cartesian coordinates, without using the Lambert $W$ function. Also, we investigate the effect of the parameter $k$ on the radial distance of the projectile showing that the radial distance oscillates from a certain critical launch angle and find an approximate expression for it. In our analysis, we consider the impact of the parameter $k$ in the kinetic energy, the potential energy, the total energy and the rate of energy loss, and in the phase space. Our results can be included in an intermediate-level classical mechanics course., Comment: 13 pages, 7 figures

Federated learning (FL) offers privacy-preserving decentralized machine learning, optimizing models at edge clients without sharing private data. Simultaneously, foundation models (FMs) have gained traction in the artificial intelligence (AI) community due to their exceptional performance across various tasks. However, integrating FMs into FL presents challenges, primarily due to their substantial size and intensive resource requirements. This is especially true when considering the resource heterogeneity in edge FL systems. We present an adaptive framework for Resource-aware Federated Foundation Models (RaFFM) to address these challenges. RaFFM introduces specialized model compression algorithms tailored for FL scenarios, such as salient parameter prioritization and high-performance subnetwork extraction. These algorithms enable dynamic scaling of given transformer-based FMs to fit heterogeneous resource constraints at the network edge during both FL's optimization and deployment stages. Experimental results demonstrate that RaFFM shows significant superiority in resource utilization efficiency and uses fewer resources to deploy FMs to FL. Despite the lower resource consumption, target models optimized by RaFFM achieve performance on par with traditional FL methods applied to full-sized FMs. This is evident across tasks in both natural language processing and computer vision domains.

We measured the spin dependence in a neutron-induced $p$-wave resonance by using a polarized epithermal neutron beam and a polarized nuclear target. Our study focuses on the 0.75~eV $p$-wave resonance state of $^{139}$La+$n$, where largely enhanced parity violation has been observed. We determined the partial neutron width of the $p$-wave resonance by measuring the spin dependence of the neutron absorption cross section between polarized $^{139}\rm{La}$ and polarized neutrons. Our findings serve as a foundation for the quantitative study of the enhancement effect of the discrete symmetry violations caused by mixing between partial amplitudes in the compound nuclei.

We use X-ray observations of quasar microlensing (sensitive to smaller compact objects than in the optical) to study the possible presence of a population of low mass black holes (from $\sim$ $10^{-3}M_{\odot}$ to $10^{-1}M_{\odot}$) in lens galaxies. We compare these observations with microlensing magnification simulations of a mixed population of stars and black holes (BHs) plus a smooth matter component. We estimate the individual mass fractions of both, stars and BHs, for three different BH masses in the range of substellar to planetary masses. Our Bayesian analysis indicates that the contribution of BHs is negligible in the substellar mass range but that a population of BHs of planetary mass (M $\lesssim$ $10^{-3}M_{\odot}$) could pass unnoticed to X-ray microlensing. We provide new upper limits to the contribution of BHs to the fraction of dark matter based on both, the quasar microlensing data in the X-ray band, and our previous estimates in the optical of intermediate-mass BHs with an additional upper limit at $M=3M_{\odot}$., Comment: 11 pages, 3 figures

Our aim is to examine the size, kinematics, and geometry of the broad-line region (BLR) in the double-lensed quasar Q 0957+561 by analyzing the impact of microlensing on various rest-frame ultraviolet broad-emission lines (BELs). We explore the influence of intrinsic variability and microlensing on the C IV, C III], and Mg II emission lines through multiple spectroscopic observations taken between April 1999 and January 2017. By utilizing the line cores as a reference for no microlensing and correcting for the long time delay between the images, we estimate the sizes of the regions emitting the broad-line wings using a Bayesian approach. Our study of the microlensing amplitudes between the lensed images of the quasar Q 0957+561 reveals differing sizes of the regions emitting the three prominent BELs C IV, C III], and Mg II. The strength of the differential microlensing indicates that the high-ionization line C IV arises from a compact inner region of the BLR with a half-light radius of $R_{1/2} \gtrsim 16.0$ lt-days, which represents a lower limit on the overall size of the BLR and is comparable to the size of the region emitting the r-band continuum in this system. A somewhat larger size of $R_{1/2}\gtrsim 44$ lt-days is obtained for the semi-forbidden line C III]. Microlensing has a weak impact on the lower-ionization line Mg II, which is emitted from a region with a half-light radius of $R_{1/2} \gtrsim 50$ lt-days. These findings suggest that the BEL regions may have distinct geometries and kinematics, with the more extended ones being spherically symmetric, and the most compact ones being nonspherical, with motions likely confined to a plane.

Foundation Models (FMs), such as LLaMA, BERT, GPT, ViT, and CLIP, have demonstrated remarkable success in a wide range of applications, driven by their ability to leverage vast amounts of data for pre-training. However, optimizing FMs often requires access to sensitive data, raising privacy concerns and limiting their applicability in many domains. In this paper, we propose the Federated Foundation Models (FFMs) paradigm, which combines the benefits of FMs and Federated Learning (FL) to enable privacy-preserving and collaborative learning across multiple end-users. We discuss the potential benefits and challenges of integrating FL into the lifespan of FMs, covering pre-training, fine-tuning, and application. We further outline potential future research avenues in FFM, including FFM pre-training, FFM fine-tuning, and federated prompt tuning, which allow the development of more personalized and context-aware models while ensuring data privacy. Moreover, we explore the possibility of continual/lifelong learning in FFMs, as increased computational power at the edge may unlock the potential for optimizing FMs using newly generated private data close to the data source. The proposed FFM concepts offer a flexible and scalable framework for training large language models in a privacy-preserving manner, setting the stage for subsequent advancements in both FM training and federated learning., Comment: Accepted at the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)

Federated Learning (FL) is extensively used to train AI/ML models in distributed and privacy-preserving settings. Participant edge devices in FL systems typically contain non-independent and identically distributed (Non-IID) private data and unevenly distributed computational resources. Preserving user data privacy while optimizing AI/ML models in a heterogeneous federated network requires us to address data and system/resource heterogeneity. To address these challenges, we propose Resource-aware Federated Learning (RaFL). RaFL allocates resource-aware specialized models to edge devices using Neural Architecture Search (NAS) and allows heterogeneous model architecture deployment by knowledge extraction and fusion. Combining NAS and FL enables on-demand customized model deployment for resource-diverse edge devices. Furthermore, we propose a multi-model architecture fusion scheme allowing the aggregation of the distributed learning results. Results demonstrate RaFL's superior resource efficiency compared to SoTA., Comment: Accepted at the 30th International European Conference on Parallel and Distributed Computing (Euro-Par 2024)

We catalog the 443 bright supernovae discovered by the All-Sky Automated Survey for Supernovae (ASAS-SN) in $2018-2020$ along with the 519 supernovae recovered by ASAS-SN and 516 additional $m_{peak}\leq18$ mag supernovae missed by ASAS-SN. Our statistical analysis focuses primarily on the 984 supernovae discovered or recovered in ASAS-SN $g$-band observations. The complete sample of 2427 ASAS-SN supernovae includes earlier $V$-band samples and unrecovered supernovae. For each supernova, we identify the host galaxy, its UV to mid-IR photometry, and the offset of the supernova from the center of the host. Updated light curves, redshifts, classifications, and host galaxy identifications supersede earlier results. With the increase of the limiting magnitude to $g\leq18$ mag, the ASAS-SN sample is roughly complete up to $m_{peak}=16.7$ mag and is $90\%$ complete for $m_{peak}\leq17.0$ mag. This is an increase from the $V$-band sample where it was roughly complete up to $m_{peak}=16.2$ mag and $70\%$ complete for $m_{peak}\leq17.0$ mag., Comment: 14 pages, 7 figures, 4 tables. Updated to reflect changes made in the published version. Tables containing the catalog data presented in this submission are included in machine-readable format as ancillary files

Neural Architecture Search (NAS) has significantly improved productivity in the design and deployment of neural networks (NN). As NAS typically evaluates multiple models by training them partially or completely, the improved productivity comes at the cost of significant carbon footprint. To alleviate this expensive training routine, zero-shot/cost proxies analyze an NN at initialization to generate a score, which correlates highly with its true accuracy. Zero-cost proxies are currently designed by experts conducting multiple cycles of empirical testing on possible algorithms, datasets, and neural architecture design spaces. This experimentation lowers productivity and is an unsustainable approach towards zero-cost proxy design as deep learning use-cases diversify in nature. Additionally, existing zero-cost proxies fail to generalize across neural architecture design spaces. In this paper, we propose a genetic programming framework to automate the discovery of zero-cost proxies for neural architecture scoring. Our methodology efficiently discovers an interpretable and generalizable zero-cost proxy that gives state of the art score-accuracy correlation on all datasets and search spaces of NASBench-201 and Network Design Spaces (NDS). We believe that this research indicates a promising direction towards automatically discovering zero-cost proxies that can work across network architecture design spaces, datasets, and tasks.

Very recently the LHCb experiment released the first measurement of the ratio $R(\Lambda_c) = {\rm BR}(\Lambda_b \to \Lambda_c\tau\bar{\nu}_\tau)/{\rm BR}(\Lambda_b \to \Lambda_c\mu\bar{\nu}_\mu)$. Moreover, the BABAR experiment reported a new result of the leptonic decay ratio of Upsilon meson $\Upsilon(3S)$, namely, $R_{\Upsilon(3S)} = {\rm BR}(\Upsilon(3S) \to \tau^+\tau^-)/{\rm BR}(\Upsilon(3S) \to \mu^+\mu^-)$. Both measurements are below their corresponding Standard Model predictions (deficit), deviating by $\sim 1.1\sigma$ and $\sim 1.8\sigma$, respectively. Moreover, the LHCb recently presented the first search of the lepton flavor violating decay $B^0 \to K^{\ast 0}\mu^\pm\tau^\mp$. Motivated by these new data, in this work we study their impact on the phenomenology of the singlet vector leptoquark ($U_1$) model addressing the hints of lepton flavor universality violation in the semileptonic decays of $B$ mesons ($B$ meson anomalies), by carrying out a global fit analysis. In general, we found that a minimal version of the $U_1$ model with a mass of 1.8 TeV can successfully explain the $B$ meson anomalies, while being compatible with all other flavor observables and LHC bounds. Interestingly, our study shows that the new observables $R(\Lambda_c)$ and $R_{\Upsilon(3S)}$ generate strong tension, leading to non-trivial effects on the global fit. Future improvements at the LHCb and Belle II experiments would help to understand their complementarity. Moreover, we also analyze the impact of the expected sensitivity on flavor observables at Belle II to provide a further test of the $U_1$ model. Finally, we study the minimal assumptions under which the $U_1$ model could, in addition, provide a combined explanation of the anomalous magnetic moment of the muon., Comment: 15 pages, 3 figures. V2: Typos fixed. New References added. Recent LHCb result on the LFV decay B^0 to K^{*0} mu^{+} tau^- (mu^{-} tau^+) included. Phenomenology analysis improved. Overall conclusions unchanged

Concerned with user data privacy, this paper presents a new federated learning (FL) method that trains machine learning models on edge devices without accessing sensitive data. Traditional FL methods, although privacy-protective, fail to manage model heterogeneity and incur high communication costs due to their reliance on aggregation methods. To address this limitation, we propose a resource-aware FL method that aggregates local knowledge from edge models and distills it into robust global knowledge through knowledge distillation. This method allows efficient multi-model knowledge fusion and the deployment of resource-aware models while preserving model heterogeneity. Our method improves communication cost and performance in heterogeneous data and models compared to existing FL algorithms. Notably, it reduces the communication cost of ResNet-32 by up to 50\% and VGG-11 by up to 10$\times$ while delivering superior performance., Comment: Accept at the 4th workshop on Artificial Intelligence and Machine Learning for Scientific Applications (AI4S), SC 23

In this work we present an analysis of production and signature of charged Higgs bosons $H_2^{\pm}$ in the version of the 3-3-1 model containing heavy leptons at the CLIC (Cern Linear Collider). The production rate is found to be significant for the direct production of $e^{-} e^{+} \rightarrow H_{2}^{+} H_{2}^{-}$. We also studied the possibility to identify it using their respective branching ratios., Comment: 16 pages, 9 Figures. arXiv admin note: substantial text overlap with arXiv:1408.5944; text overlap with arXiv:1311.0845, arXiv:1205.4042

We have built a new model for the lens system SDSS J1004+4112 including the recently measured time delay of the fourth quasar image. This time delay has a strong influence on the inner mass distribution of the lensing cluster ($\rho \propto r^{-\alpha}$) allowing us to determine $\alpha=1.18^{+0.02(+0.11)}_{-0.03(-0.18)}$ at the 68% (95%) confidence level in agreement with hydrodynamical simulations of massive galaxy clusters. We find an offset between the brightest cluster galaxy (BCG) and the dark matter halo of $3.8^{+0.6(+1.4)}_{-0.7(-1.3)}$ kpc at 68% (95%) confidence which is compatible with other galaxy cluster measurements. As an observational challenge, the estimated time delay between the leading image C and the faint (I=24.7) fifth image E is roughly 8 yr., Comment: 9 pages, 3 figures, 4 tables. ApJ in press

We present new light curves for the four bright images of the five image cluster-lensed quasar gravitational lens system SDSS~J1004+4112. The light curves span 14.5 yr and allow measurement of the time delay between the trailing bright quasar image D and the leading image C. When we fit all four light curves simultaneously and combine the models using the Bayes information criterion, we find a time delay of $\Delta t_{DC}= 2458.47 \pm 1.02$ days (6.73 yr), the longest ever measured for a gravitational lens. For the other two independent time delays we obtain $\Delta t_{BC}=782.20 \pm 0.43$ days (2.14 yr) and $\Delta t_{AC}= 825.23 \pm 0.46$ days (2.26 yr), in agreement with previous results. The information criterion is needed to weight the results for light curve models with different polynomial orders for the intrinsic variability and the effects of differential microlensing. The results using the Akaike information criterion are slightly different, but, in practice, the absolute delay errors are all dominated by the $\sim 4\%$ cosmic variance in the delays rather than the statistical or systematic measurement uncertainties. Despite the lens being a cluster, the quasar images show slow differential variability due to microlensing at the level of a few tenths of a magnitude., Comment: 12 pages, 5 figures. ApJ in press