Search

Continual learning is a subfield of machine learning, which aims to allow machine learning models to continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take a step back, and ask: "Why should one care about continual learning in the first place?". We set the stage by examining recent continual learning papers published at four major machine learning conferences, and show that memory-constrained settings dominate the field. Then, we discuss five open problems in machine learning, and even though they might seem unrelated to continual learning at first sight, we show that continual learning will inevitably be part of their solution. These problems are model editing, personalization and specialization, on-device learning, faster (re-)training and reinforcement learning. Finally, by comparing the desiderata from these unsolved problems and the current assumptions in continual learning, we highlight and discuss four future directions for continual learning research. We hope that this work offers an interesting perspective on the future of continual learning, while displaying its potential value and the paths we have to pursue in order to make it successful. This work is the result of the many discussions the authors had at the Dagstuhl seminar on Deep Continual Learning, in March 2023.

Deep learning models have demonstrated remarkable results for various computer vision tasks, including the realm of medical imaging. However, their application in the medical domain is limited due to the requirement for large amounts of training data, which can be both challenging and expensive to obtain. To mitigate this, pre-trained models have been fine-tuned on domain-specific data, but such an approach can suffer from inductive biases. Furthermore, deep learning models struggle to learn the relationship between spatially distant features and their importance, as convolution operations treat all pixels equally. Pioneering a novel solution to this challenge, we employ the Image Foresting Transform to optimally segment images into superpixels. These superpixels are subsequently transformed into graph-structured data, enabling the proficient extraction of features and modeling of relationships using Graph Neural Networks (GNNs). Our method harnesses an ensemble of three distinct GNN architectures to boost its robustness. In our evaluations targeting pneumonia classification, our methodology surpassed prevailing Deep Neural Networks (DNNs) in performance, all while drastically cutting down on the parameter count. This not only trims down the expenses tied to data but also accelerates training and minimizes bias. Consequently, our proposition offers a sturdy, economically viable, and scalable strategy for medical image classification, significantly diminishing dependency on extensive training data sets., Comment: Our code is available at https://github.com/aryan-at-ul/AICS_2023_submission

Recent years have seen considerable progress in the continual training of deep neural networks, predominantly thanks to approaches that add replay or regularization terms to the loss function to approximate the joint loss over all tasks so far. However, we show that even with a perfect approximation to the joint loss, these approaches still suffer from temporary but substantial forgetting when starting to train on a new task. Motivated by this 'stability gap', we propose that continual learning strategies should focus not only on the optimization objective, but also on the way this objective is optimized. While there is some continual learning work that alters the optimization trajectory (e.g., using gradient projection techniques), this line of research is positioned as alternative to improving the optimization objective, while we argue it should be complementary. In search of empirical support for our proposition, we perform a series of pre-registered experiments combining replay-approximated joint objectives with gradient projection-based optimization routines. However, this first experimental attempt fails to show clear and consistent benefits. Nevertheless, our conceptual arguments, as well as some of our empirical results, demonstrate the distinctive importance of the optimization trajectory in continual learning, thereby opening up a new direction for continual learning research., Comment: Full paper version of pre-registered report accepted at the 1st ContinualAI Unconference. The originally submitted pre-registered proposal can be found at arXiv:2311.04898v1

Machine learning has rapidly become a tool of choice for the astronomical community. It is being applied across a wide range of wavelengths and problems, from the classification of transients to neural network emulators of cosmological simulations, and is shifting paradigms about how we generate and report scientific results. At the same time, this class of method comes with its own set of best practices, challenges, and drawbacks, which, at present, are often reported on incompletely in the astrophysical literature. With this paper, we aim to provide a primer to the astronomical community, including authors, reviewers, and editors, on how to implement machine learning models and report their results in a way that ensures the accuracy of the results, reproducibility of the findings, and usefulness of the method., Comment: 14 pages, 3 figures; submitted to the Bulletin of the American Astronomical Society

The goal of this work is to probe the total mass distribution of early-type galaxies with globular clusters (GCs) as kinematic tracers, by constraining the parameters of the profile with a flexible modelling approach. To that end, we leverage the extended spatial distribution of GCs from the SLUGGS survey ($\langle R_{\rm GC,\ max} \rangle \sim 8R_{\rm e}$) in combination with discrete dynamical modelling. We use discrete Jeans anisotropic modelling in cylindrical coordinates to determine the velocity moments at the location of the GCs in our sample. We use a Bayesian framework to determine the best-fit parameters of the total mass density profile and orbital properties of the GC systems. We find that the orbital properties (anisotropy and rotation of the dispersion-dominated GC systems) minimally impact the measurements of the inner slope and enclosed mass, while a strong presence of dynamically-distinct subpopulations or low numbers of kinematic tracers can bias the results. Owing to the large spatial extent of the tracers our method is sensitive to the intrinsic inner slope of the total mass profile and we find $\overline{\alpha} = -1.88\pm 0.01$ for 12 galaxies with robust measurements. To compare our results with literature values we fit a single power-law profile to the resulting total mass density. In the radial range 0.1-4~$R_{\rm e}$ our measured slope has a value of $\langle \gamma_{\rm tot}\rangle = -2.22\pm0.14$ and is in good agreement with the literature., Comment: 17 pages, 13 Figures, 8 Tables, Accepted for publication in A&A

We employ a recently developed population-orbit superposition technique to simultaneously fit the stellar kinematic and age maps of 82 CALIFA spiral galaxies and obtain the ages of stars in different dynamical structures. We first evaluated the capabilities of this method on CALIFA-like mock data created from the Auriga simulations. The recovered mean ages of dynamically cold, warm, and hot components match the true values well, with an observational error of up to $20\%$ in the mock age maps. For CALIFA spiral galaxies, we find that the stellar ages of the cold, warm, and hot components all increase with the stellar mass of the galaxies, from $\overline{t_{\rm cold}}\sim2.2$ Gyr, $\overline{t_{\rm warm}}\sim2.3$ Gyr, and $\overline{t_{\rm hot}}\sim2.6$ Gyr for galaxies with stellar mass $M_*<10^{10}\,\rm M_{\odot}$, to $\overline{t_{\rm cold}}\sim4.0$ Gyr, $\overline{t_{\rm warm}}\sim5.1$ Gyr, and $\overline{t_{\rm hot}}\sim5.9$ Gyr for galaxies with $M_*>10^{11}\,\rm M_{\odot}$. About $80\%$ of the galaxies in our sample have $t_{\rm hot}>t_{\rm cold}$, and the mean values of $t_{\rm hot}-t_{\rm cold}$ also increase with stellar mass, from $0.7_{-0.2}^{+0.6}$ Gyr in low-mass galaxies ($10^{8.9}\,\rm M_{\odot}

The merger history of a galaxy is thought to be one of the major factors determining its internal dynamics, with galaxies having undergone different types or mergers (e.g. dry, minor or major mergers) predicted to show different dynamical properties. We study the instantaneous orbital distribution of galaxies in the Eagle simulation, colouring the orbits of the stellar particles by their stellar age, in order to understand whether stars form in particular orbits (e.g. in a thin or thick disc). We first show that Eagle reproduces well the observed stellar mass fractions in different stellar orbital families as a function of stellar mass and spin parameter at z = 0. We find that the youngest stars reside in a thin disc component that can extend to the very inner regions of galaxies, and that older stars have warmer orbits, with the oldest ones showing orbits consistent with both hot and counter-rotating classifications, which is consistent with the trend found in the Milky-Way and other disc galaxies. We also show that counter-rotating orbits trace galaxy mergers - in particular dry mergers, and that in the absence of mergers, counter-rotating orbits can also be born from highly misaligned gas accretion that leads to star formation., Comment: Submitted to MNRAS, 21 pages

We apply the barred Schwarzschild method developed by Tahmasebzadeh et al. (2022) to a barred S0 galaxy, NGC 4371, observed by IFU instruments from the TIMER and ATLAS3D projects. We construct the gravitational potential by combining a fixed black hole mass, a spherical dark matter halo, and stellar mass distribution deprojected from $3.6$ $\mu$m S$^4$G image considering an axisymmetric disk and a triaxial bar. We independently modelled kinematic data from TIMER and ATLAS3D. Both models fit the data remarkably well. We find a consistent bar pattern speed from the two sets of models with $\Omega_{\rm p} = 23.6 \pm 2.8 \hspace{.08cm} \mathrm{km \hspace{.04cm} s^{-1} \hspace{.04cm} kpc^{-1} }$ and $\Omega_{\rm p} = 22.4 \pm 3.5 \hspace{.08cm} \mathrm{km \hspace{.04cm} s^{-1} \hspace{.04cm} kpc^{-1} }$, respectively. The dimensionless bar rotation parameter is determined to be $ 1.88 \pm 0.37$, indicating a likely slow bar in NGC 4371. Additionally, our model predicts a high amount of dark matter within the bar region ($M_{\rm DM}/ M_{\rm total}$ $\sim 0.51 \pm 0.06$), which, aligned with the predictions of cosmological simulations, indicates that fast bars are generally found in baryon-dominated disks. Based on the best-fitting model, we further decompose the galaxy into multiple 3D orbital structures, including a BP/X bar, a classical bulge, a nuclear disk, and a main disk. The BP/X bar is not perfectly included in the input 3D density model, but BP/X-supporting orbits are picked through the fitting to the kinematic data. This is the first time a real barred galaxy has been modelled utilizing the Schwarzschild method including a 3D bar., Comment: 25 pages, 19 figures, Accepted for publication in MNRAS

The mass-metallicity relation (MZR) represents one of the most important scaling relations in the context of galaxy evolution, comprising a positive correlation between stellar mass and metallicity (Z). The fundamental metallicity relation (FMR) introduces a new parameter, the star formation rate (SFR), in the dependence. While several studies found that Z is anti-correlated with the SFR at fixed mass, the validity of this statement has been questioned extensively and no widely-accepted consensus has been reached yet. With this work, we investigate the FMR in nine nearby, spatially-resolved, dwarf galaxies, using gas diagnostics on integral-field spectroscopic data of the Multi Unit Spectroscopic Explorer (MUSE), pushing such investigations to lower galaxy masses and higher resolutions. We find that both the MZR and FMR exhibit different behaviours within different star forming regions of the galaxies. We find that the SFR surface density - metallicity anti-correlation is tighter in the low-mass galaxies of our sample. For all the galaxies considered, we find a SFR surface density - stellar mass surface density correlation. We propose that the main reason behind these findings is connected to the accretion mechanisms of the gas fuelling star formation -- low-mass, metal-poor galaxies accrete pristine gas from the intergalactic medium, while in more massive and metal-enriched systems the gas responsible for star formation is recycled from previous star forming episodes., Comment: 15 pages, 8 figures, accepted for publication in Astronomy & Astrophysics

We construct C*-dynamical systems for the dynamics of classical infinite particle systems describing harmonic oscillators interacting with arbitrarily many neighbors on lattices, as well on more general structures. Our approach allows particles with varying masses, varying frequencies, irregularly placed lattice sites and varying interactions subject to a simple summability constraint. A key role is played by the commutative resolvent algebra, which is a C*-algebra of bounded continuous functions on an infinite dimensional vector space, and in a strong sense the classical limit of the Buchholz--Grundling resolvent algebra, which suggests that quantum analogs of our results are likely to exist. For a general class of Hamiltonians, we show that the commutative resolvent algebra is time-stable, and admits a time-stable sub-algebra on which the dynamics is strongly continuous, therefore obtaining a C*-dynamical system., Comment: 40 pages

We present an observational study of the dark matter fraction in star-forming disk-like galaxies up to redshift $z \sim 2.5$, selected from publicly available integral field spectroscropic surveys: KMOS$^{\rm 3D}$, KGES, and KROSS. To model the $H\alpha$ kinematics of these galaxies, we employ 3D forward-modelling, which incorporates beam-smearing and inclination corrections, and yields rotation curves. Subsequently, these rotation curves are corrected for gas pressure gradients, resulting in circular velocity curves or `intrinsic' rotation curves. Our final sample comprises of 263 rotationally supported main sequence star-forming galaxies with redshifts ranging from $0.6 \leq z < 2.5$. We estimate the dark matter fraction of these galaxies by subtracting the baryonic mass from the total mass, where the total mass is derived from the intrinsic rotation curves. We provide novel observational evidence, suggesting that at a fixed redshift, the dark matter fraction gradually increases with radius such that the outskirts of galaxies are dark matter dominated, similarly to local star-forming disk galaxies. This observed dark matter fraction exhibits a decreasing trend with increasing redshift and, on average, the fraction within the effective radius (upto outskirts) remains above 50\%, similar to locals. We investigate the relationships between dark matter, baryon surface density, and circular velocity of galaxies. We observe that low stellar mass galaxies, with $\log(M_{\rm star}\ [\mathrm{M_\odot}]) \leq 10.0$, undergo a higher degree of evolution, which may be attributed to the hierarchical merging of galaxies. Most importantly, we discuss several sources of uncertainties and current limitations in the field, as well as their impact on the measurements of dark matter fraction and its trend across galactic scales and cosmic time., Comment: Comments are welcome

Omega Centauri ($\omega$ Cen) is the most massive globular cluster of the Milky Way and has been the focus of many studies that reveal the complexity of its stellar populations and kinematics. However, most previous studies have used photometric and spectroscopic datasets with limited spatial or magnitude coverage, while we aim to investigate it having full spatial coverage out to its half-light radius and stars ranging from the main sequence to the tip of the red giant branch. This is the first paper in a new survey of $\omega$ Cen that combines uniform imaging and spectroscopic data out to its half-light radius to study its stellar populations, kinematics, and formation history. In this paper, we present an unprecedented MUSE spectroscopic dataset combining 87 new MUSE pointings with previous observations collected from guaranteed time observations. We extract spectra of more than 300,000 stars reaching more than two magnitudes below the main sequence turn-off. We use these spectra to derive metallicity and line-of-sight velocity measurements and determine robust uncertainties on these quantities using repeat measurements. Applying quality cuts we achieve signal-to-noise ratios of 16.47/73.51 and mean metallicity errors of 0.174/0.031 dex for the main sequence stars (18 mag $\rm < mag_{F625W}<$22 mag) and red giant branch stars (16 mag $<\rm mag_{F625W}<$10 mag), respectively. We correct the metallicities for atomic diffusion and identify foreground stars. This massive spectroscopic dataset will enable future studies that will transform our understanding of $\omega$ Cen, allowing us to investigate the stellar populations, ages, and kinematics in great detail., Comment: 27 pages, 18 figures, 3 tables, published in ApJ, the catalog is available in the online material of the published article; typos corrected in this version

Using MUSE data, we investigate the radial gradients of stellar population properties (namely age, [M/H], and the abundance ratio of $\alpha$ elements [$\alpha$/Fe]) for a sample of nine dwarf early-type (dE) galaxies with log(M$_{\star}$/M$_{\odot}$) $\sim$ 9.0 and an infall time onto the Virgo cluster of 2-3Gyr ago. We followed a similar approach as in Bidaran et al. (2022) to derive their stellar population properties and star formation histories (SFHs) through fitting observed spectral indices and full spectral fitting, respectively. We find that these nine dE galaxies have truncated [Mg/Fe]vs.[Fe/H] profiles than equally-massive Virgo dE galaxies with longer past infall times. Short profiles of three dE galaxies are the result of their intense star formation which has been quenched long before their accretion onto the Virgo cluster, possibly as a result of their group environment. In the remaining six dE galaxies, profiles mainly trace a recent episode of star burst within 0.4R$_{\rm e}$ which results in higher light-weighted [$\alpha$/Fe] values. The latter SFH peak can be due to ram pressure exerted by the Virgo cluster at the time of the accretion of the dE galaxies. Also, we show that younger, more metal-rich and less $\alpha$-enhanced stellar populations dominate their inner regions (i.e., < 0.4R$_{\rm e}$) resulting in mainly flat $\nabla_{\rm age}$, negative $\nabla_{\rm [M/H]}$ and positive $\nabla_{\rm [\alpha/Fe]}$. We find that with increasing log($\sigma_{\rm Re}$) of dE galaxies, $\nabla_{\rm age}$ and $\nabla_{\rm [\alpha/Fe]}$ flatten, and the latter correlation persists even after including early-type galaxies up to log($\sigma_{\rm Re}$ $\sim$ 2.5), possibly due to the more extended star formation activity in the inner regions of dEs, as opposed to more massive early-type galaxies., Comment: 18 pages, 8 figures, accepted in MNRAS

When galactic discs settle and become massive enough, they are able to form stellar bars. These non-axisymmetric structures induce shocks in the gas, causing it to flow to the centre where nuclear structures, such as nuclear discs and rings, are formed. Previous theoretical and observational studies have hinted at the co-evolution of bars and nuclear discs, suggesting that nuclear discs grow "inside-out", thereby proposing that smaller discs live in younger bars. Nevertheless, it remains unclear how the bar and the nuclear structures form and evolve with time. The smallest nuclear discs discovered to date tend to be larger than $\sim200~\rm{pc}$, even though some theoretical studies find that when nuclear discs form they can be much smaller. Using MUSE archival data, we report for the first time two extragalactic nuclear discs with radius sizes below $100~\rm{pc}$. Additionally, our estimations reveal the youngest bars found to date. We estimate that the bars in these galaxies formed $4.50^{+1.60}_{-1.10}\rm{(sys)}^{+1.00}_{-0.75}\rm{(stat)}$ and $0.7^{+2.60}\rm{(sys)}^{+0.05}_{-0.05}\rm{(stat)}~\rm{Gyr}$ ago, for NGC\,289 and NGC\,1566, respectively. This suggests that at least some disc galaxies in the Local Universe may still be dynamically settling. By adding these results to previous findings in the literature, we retrieve a stronger correlation between nuclear disc size and bar length and we derive a tentative exponential growth scenario for nuclear discs., Comment: Accepted in A&A (in press), 13 pages, 5 figures, 2 tabels

Graph-based neural network models are gaining traction in the field of representation learning due to their ability to uncover latent topological relationships between entities that are otherwise challenging to identify. These models have been employed across a diverse range of domains, encompassing drug discovery, protein interactions, semantic segmentation, and fluid dynamics research. In this study, we investigate the potential of Graph Neural Networks (GNNs) for medical image classification. We introduce a novel model that combines GNNs and edge convolution, leveraging the interconnectedness of RGB channel feature values to strongly represent connections between crucial graph nodes. Our proposed model not only performs on par with state-of-the-art Deep Neural Networks (DNNs) but does so with 1000 times fewer parameters, resulting in reduced training time and data requirements. We compare our Graph Convolutional Neural Network (GCNN) to pre-trained DNNs for classifying MedMNIST dataset classes, revealing promising prospects for GNNs in medical image analysis. Our results also encourage further exploration of advanced graph-based models such as Graph Attention Networks (GAT) and Graph Auto-Encoders in the medical imaging domain. The proposed model yields more reliable, interpretable, and accurate outcomes for tasks like semantic segmentation and image classification compared to simpler GCNNs

We use CEERS JWST/NIRCam imaging to measure rest-frame near-IR light profiles of $>$500 $M_\star>10^{10}~M_\odot$ galaxies in the redshift range $0.510^{11}~M_\odot$) quiescent galaxies between $z=0.5$ and $z=2.3$, again comparing $R_{\rm{opt}}$, \rmass~and \rmassd. We conclude that the main tenets of the size evolution narrative established over the past 20 years, based on rest-frame optical light profile analysis, still hold in the era of JWST/NIRCam observations in the rest-frame near-IR., Comment: Submitted to ApJ. Comments welcome

We present the first results from the Revealing Low-Luminosity Active Galactic Nuclei (ReveaLLAGN) survey, a JWST survey of seven nearby LLAGN. We focus on two observations with the Mid-Infrared Instrument's (MIRI) Medium Resolution Spectrograph (MRS) of the nuclei of NGC 1052 and Sombrero (NGC 4594 / M104). We also compare these data to public JWST data of a higher-luminosity AGN, NGC 7319 and NGC 7469. JWST clearly separates the AGN spectrum from the galaxy light even in Sombrero, the faintest target in our survey; the AGN components have very red spectra. We find that the emission-line widths in both NGC 1052 and Sombrero increase with increasing ionization potential, with FWHM > 1000 km/s for lines with ionization potential > 50 eV. These lines are also significantly blue-shifted in both LLAGN. The high ionization potential lines in NGC 7319 show neither broad widths or significant blue shifts. Many of the lower ionization potential emission lines in Sombrero show significant blue wings extending > 1000 km/s. These features and the emission-line maps in both galaxies are consistent with outflows along the jet direction. Sombrero has the lowest luminosity high-ionization potential lines ([Ne V] and [O IV]) ever measured in the mid-IR, but the relative strengths of these lines are consistent with higher luminosity AGN. On the other hand, the [Ne V] emission is much weaker relative to the [Ne III] and [Ne II] lines of higher-luminosity AGN. These initial results show the great promise that JWST holds for identifying and studying the physical nature of LLAGN., Comment: Accepted to ApJ Feb 28, 2024

Using VLT/MUSE integral-field spectroscopic data for the barred spiral galaxy NGC 1097, we explore techniques that can be used to search for extended coherent shocks that can drive gas inflows in centres of galaxies. Such shocks should appear as coherent velocity jumps in gas kinematic maps, but this appearance can be distorted by inaccurate extraction of the velocity values and dominated by the global rotational flow and local perturbations like stellar outflows. We include multiple components in the emission-line fits, which corrects the extracted velocity values and reveals emission associated with AGN outflows. We show that removal of the global rotational flow by subtracting the circular velocity of a fitted flat disk can produce artefacts that obscure signatures of the shocks in the residual velocities if the inner part of the disk is warped or if gas is moving around the centre on elongated (non-circular) trajectories. As an alternative, we propose a model-independent method which examines differences in the LOSVD moments of H$\alpha$ and [N II]$\lambda$6583. This new method successfully reveals the presence of continuous shocks in the regions inward from the nuclear ring of NGC 1097, in agreement with nuclear spiral models., Comment: MNRAS accepted (19 pages, 18 figures, 1 table)

Class-incremental learning (CIL) is a particularly challenging variant of continual learning, where the goal is to learn to discriminate between all classes presented in an incremental fashion. Existing approaches often suffer from excessive forgetting and imbalance of the scores assigned to classes that have not been seen together during training. In this study, we introduce a novel approach, Prediction Error-based Classification (PEC), which differs from traditional discriminative and generative classification paradigms. PEC computes a class score by measuring the prediction error of a model trained to replicate the outputs of a frozen random neural network on data from that class. The method can be interpreted as approximating a classification rule based on Gaussian Process posterior variance. PEC offers several practical advantages, including sample efficiency, ease of tuning, and effectiveness even when data are presented one class at a time. Our empirical results show that PEC performs strongly in single-pass-through-data CIL, outperforming other rehearsal-free baselines in all cases and rehearsal-based methods with moderate replay buffer size in most cases across multiple benchmarks., Comment: ICLR 2024 camera ready

Cosmological simulations predict dark matter shapes that deviate from spherical symmetry. The exact shape depends on the prescription of the simulation and the interplay between dark matter and baryons. This signature is most pronounced in the diffuse galactic haloes that can be observationally probed with planetary nebulae and globular clusters (GCs). The kinematic observations of these halo tracers support intrinsic triaxial shape for the mass generating the gravitational potential. With discrete axisymmetric modelling of GCs as the halo tracers of NGC 5128 we investigate the overall mass distribution of this nearby giant elliptical galaxy. Our modelling approach constrains $c_{200}$, $(M/L)_{\star, B}$ and inclination. We derive a preliminary $M_{200}\sim 1 \times 10^{12}$ M$_\odot$ and flattening $q_{\mathrm{DM}}\sim 1.3$ indicative of prolate/triaxial halo for NGC 5128., Comment: 6 pages, 2 figures, Accepted for publication in Proceedings of the IAUS 379: Dynamical Masses of Local Group Galaxies, Potsdam, March 20-24, 2023

Triaxial dynamical models of massive galaxies observed in the ATLAS3D project can provide new insights into the complex evolutionary processes that shape galaxies. The ATLAS3D survey is ideal as the sample comprises a good mix of fast and slow rotators with vastly different mass assembly histories. We present a detailed dynamical study with our triaxial modelling code DYNAMITE, which models galaxies as a superposition of their stellar orbits. The models allow us to constrain the intrinsic shape of the stellar component, the distributions of the visible and invisible matter and the orbit distribution in these nearby early-type galaxies and to relate it with different evolutionary scenarios. Triaxial modelling is essential for these galaxies to understand their complex kinematical features., Comment: 4 pages, 3 figures, Proceeding of IAU Symposium 379: Dynamical Masses of Local Group galaxies, ed. P. Bonifacio, M.-R. Cioni, F. Hammer, M. Pawlowski, and S. Taibi

We study the evolutionary path of the Fornax cluster galaxy NGC$~$1436, which is known to be currently transitioning from a spiral into a lenticular morphology. This galaxy hosts an inner star-forming disc and an outer quiescent disc, and we analyse data from the MeerKAT Fornax Survey, ALMA, and the Fornax3D survey to study the interstellar medium and the stellar populations of both disc components. Thanks to the combination of high resolution and sensitivity of the MeerKAT data, we find that the $\textrm{H}\scriptstyle\mathrm{I}$ is entirely confined within the inner star-forming disc, and that its kinematics is coincident with that of the CO. The cold gas disc is now well settled, which suggests that the galaxy has not been affected by any environmental interactions in the last $\sim1~$Gyr. The star formation history derived from the Fornax3D data shows that both the inner and outer disc experienced a burst of star formation $\sim5$ Gyr ago, followed by rapid quenching in the outer disc and by slow quenching in the inner disc, which continues forming stars to this day. We claim that NGC$~$1436 has begun to effectively interact with the cluster environment 5$~$Gyr ago, when a combination of gravitational and hydrodynamical interactions caused the temporary enhancement of the star-formation rate. Furthermore, due to the weaker gravitational binding $\textrm{H}\scriptstyle\mathrm{I}$ was stripped from the outer disc, causing its rapid quenching. At the same time, accretion of gas onto the inner disc stopped, causing slow quenching in this region., Comment: Monthly Notices of the Royal Astronomical Society (MNRAS), accepted for publication. Data available at the MeerKAT Fornax Survey website https://sites.google.com/inaf.it/meerkatfornaxsurvey

Supermassive black holes (SMBHs) and nuclear star clusters (NSCs) co-exist in many galaxies. While the formation history of the black hole is essentially lost, NSCs preserve their evolutionary history imprinted onto their stellar populations and kinematics. Studying SMBHs and NSCs in tandem might help us to ultimately reveal the build-up of galaxy centres. In this study, we combine large-scale VLT/MUSE and high-resolution adaptive-optics-assisted VLT/SINFONI observations of the early-type galaxy FCC 47 with the goal being to assess the effect of a spatially (non-)variable initial mass function (IMF) on the determination of the mass of the putative SMBH in this galaxy. We achieve this by performing DYNAMITE Schwarzschild orbit-superposition modelling of the galaxy and its NSC. In order to properly take account of the stellar mass contribution to the galaxy potential, we create mass maps using a varying stellar mass-to-light ratio derived from single stellar population models with fixed and with spatially varying IMFs. Using the two mass maps, we estimate black hole masses of $(7.1^{+0.8}_{-1.1})\times 10^7\,M_{\odot}$ and $(4.4^{+1.2}_{-2.1}) \times 10^7\,M_{\odot}$ at $3\sigma$ signifance, respectively. Compared to models with constant stellar-mass-to-light ratio, the black hole masses decrease by 15% and 48%, respectively. Therefore, a varying IMF, both in its functional form and spatially across the galaxy, has a non-negligible effect on the SMBH mass estimate. Furthermore, we find that the SMBH in FCC 47 has probably not grown over-massive compared to its very over-massive NSC., Comment: 23 pages 19 Figures, accepted for publication in A&A

The Dobrushin-Lanford-Ruelle (DLR) condition and the classical Kubo-Martin-Schwinger (KMS) condition are considered in the context of classical lattice systems. In particular, we prove that these conditions are equivalent for the case of a lattice spin system with values in a compact symplectic manifold by showing that infinite volume Gibbs states are in bijection with KMS states., Comment: 13 pages

Continual learning research has shown that neural networks suffer from catastrophic forgetting "at the output level", but it is debated whether this is also the case at the level of learned representations. Multiple recent studies ascribe representations a certain level of innate robustness against forgetting -- that they only forget minimally in comparison with forgetting at the output level. We revisit and expand upon the experiments that revealed this difference in forgetting and illustrate the coexistence of two phenomena that affect the quality of continually learned representations: knowledge accumulation and feature forgetting. Taking both aspects into account, we show that, even though forgetting in the representation (i.e. feature forgetting) can be small in absolute terms, when measuring relative to how much was learned during a task, forgetting in the representation tends to be just as catastrophic as forgetting at the output level. Next we show that this feature forgetting is problematic as it substantially slows down the incremental learning of good general representations (i.e. knowledge accumulation). Finally, we study how feature forgetting and knowledge accumulation are affected by different types of continual learning methods., Comment: TMLR 2024

Using very deep, high spectral resolution data from the SAMI Integral Field Spectrograph we study the stellar population properties of a sample of dwarf galaxies in the Fornax Cluster, down to a stellar mass of $10^{7}$ M$_{\odot}$, which has never been done outside the Local Group. We use full spectral fitting to obtain stellar population parameters. Adding massive galaxies from the ATLAS$^{3D}$ project, which we re-analysed, and the satellite galaxies of the Milky Way, we obtained a galaxy sample that covers the stellar mass range $10^{4}$ to $10^{12} M_{\odot}$. Using this large range we find that the mass - metallicity relation is not linear. We also find that the [$\alpha$/Fe]-stellar mass relation of the full sample shows a U-shape, with a minimum in [$\alpha$/Fe] for masses between $10^{9}-10^{10} M_{\odot}$. The relation between [$\alpha$/Fe] and stellar mass can be understood in the following way: When the faintest galaxies enter the cluster environment, a rapid burst of star formation is induced, after which the gas content is blown away by various quenching mechanisms. This fast star formation causes high [$\alpha$/Fe] values, like in the Galactic halo. More massive galaxies will manage to keep their gas longer and form several bursts of star formation, with lower [$\alpha$/Fe] as a result. For massive galaxies, stellar populations are regulated by internal processes, leading to [$\alpha$/Fe] increasing with mass. We confirm this model by showing that [$\alpha$/Fe] correlates with clustercentric distance in three nearby clusters, and also in the halo of the Milky Way., Comment: Accepted for publication in MNRAS. (17 pages, 11 figures)

Most dynamical models of galaxies to date assume axisymmetry, which is not representative of a significant fraction of massive galaxies. We have built triaxial orbit-superposition Schwarzschild models of galaxies observed by the SAMI Galaxy Survey, in order to reconstruct their inner orbital structure and mass distribution. The sample consists of 153 passive galaxies with total stellar masses in the range $10^{9.5}$ to $10^{12} M_{\odot}$. We present an analysis of the internal structures and intrinsic properties of these galaxies as a function of their environment. We measure their environment using three proxies: central or satellite designation, halo mass and local $5^{th}$ nearest neighbour galaxy density. We find that although these intrinsic properties correlate most strongly with stellar mass, environment does play a secondary role: at fixed stellar mass, galaxies in the densest regions are more radially anisotropic. In addition, central galaxies, and galaxies in high local densities show lower values of edge-on spin parameter proxy \lam. We also find suggestions of a possible trend of the fractions of orbits with environment for lower-mass galaxies (between $10^{9.5}$ and $10^{11} M_{\odot}$) such that, at fixed stellar mass, galaxies in higher local densities and halo mass have higher fractions of hot orbits and lower fractions of warm orbits. Our results demonstrate that after stellar mass, environment does play a role in shaping present-day passive galaxies., Comment: 21 pages. Accepted for publication in MNRAS

We present a unified approach to (bi-)orthogonal basis sets for gravitating systems. Central to our discussion is the notion of mutual gravitational energy, which gives rise to the self-energy inner product on mass densities. We consider a first-order differential operator that is self-adjoint with respect to this inner product, and prove a general theorem that gives the conditions under which a (bi-)orthogonal basis set arises by repeated application of this differential operator. We then show that these conditions are fulfilled by all the families of analytical basis sets with infinite extent that have been discovered to date. The new theoretical framework turns out to be closely connected to Fourier-Mellin transforms, and it is a powerful tool for constructing general basis sets. We demonstrate this by deriving a basis set for the isochrone model and demonstrating its numerical reliability by reproducing a known result concerning unstable radial modes., Comment: to be published in Astronomy & Astrophysics

We trace the evolution of the edge-on spiral galaxy NGC 3501, making use of its stellar populations extracted from deep integral-field spectroscopy MUSE observations. We present stellar kinematic and population maps, as well as the star formation history, of the south-western half of the galaxy. The derived maps of the stellar line-of-sight velocity and velocity dispersion are quite regular, show disc-like rotation, and no other structural component of the galaxy. However, maps of the stellar populations exhibit structures in the mass-weighted and light-weighted age, total metallicity and [Mg/Fe] abundance. These maps indicate that NGC 3501 is a young galaxy, consisting mostly of stars with ages between 2 to 8 Gyr. Also, they show a thicker more extended structure that is metal-poor and $\alpha$-rich, and another inner metal-rich and $\alpha$-poor one with smaller radial extension. While previous studies revealed that NGC 3501 shows only one morphological disc component in its vertical structure, we divided the galaxy into two regions: an inner metal-rich midplane and a metal-poor thicker envelope. Comparing the star formation history of the inner thinner metal-rich disc and the thicker metal-poor disc, we see that the metal-rich component evolved more steadily, while the metal-poor one experienced several bursts of star formation. We propose this spiral galaxy is being observed in an early evolutionary phase, with a thicker disc already in place and an inner thin disc in an early formation stage. So we are probably witnessing the birth of a future massive thin disc, continuously growing embedded in a preexisting thicker disc., Comment: 15 pages, 12 figures. Accepted for publication in MNRAS

We apply a population-orbit superposition method to 16 galaxies in the Fornax cluster observed with MUSE/VLT in the context of the Fornax3D project. By fitting the luminosity distribution, stellar kinematics, and age and metallicity maps simultaneously, we obtained the internal stellar orbit distribution, as well as the age and metallicity distribution of stars on different orbits for each galaxy. Based on the model, we decompose each galaxy into a dynamically cold disk (orbital circularity $\lambda_z\ge0.8$) and a dynamically hot non-disk component (orbital circularity $\lambda_z<0.8$), and obtain the surface-brightness, age, and metallicity radial profiles of each component. The galaxy infall time into the cluster is strongly correlated with galaxy cold-disk age with older cold disks in ancient infallers. We quantify the infall time $t_{\rm infall}$ of each galaxy with its cold-disk age using a correlation calibrated with TNG50 cosmological simulations. For galaxies in the Fornax cluster, we found that the luminosity fraction of cold disk in galaxies with $t_{\rm infall}>8$ Gyr are a factor of $\sim 4$ lower than in more recent infallers while controlling for total stellar mass. Nine of the 16 galaxies have spatially extended cold disks, and most of them show positive or zero age gradients; stars in the inner disk are $\sim 2-5$ Gyr younger than that in the outer disk, in contrast to the expectation of inside-out growth. Our results indicate that the assembly of cold disks in galaxies is strongly affected by their infall into clusters, by either removal of gas in outer regions or even tidally stripping or heating part of the pre-existing disks. Star formation in outer disks can stop quickly after the galaxy falls into the cluster, while star formation in the inner disks can last for a few Gyrs more, building the positive age gradient measured in cold disks., Comment: 31 pages, 32 figures, Accepted to A&A

The epoch in which galactic discs settle is a major benchmark to test models of galaxy formation and evolution but is as yet largely unknown. Once discs settle and become self-gravitating enough, stellar bars are able to form; therefore, determining the ages of bars can shed light on the epoch of disc settling, and on the onset of secular evolution. Nevertheless, until now, timing when the bar formed has proven challenging. In this work, we present a new methodology for obtaining the bar age, using the star formation history of nuclear discs. Nuclear discs are rotation-supported structures, built by gas pushed to the centre via bar-induced torques, and their formation is thus coincident with bar formation. In particular, we use integral field spectroscopic (IFS) data from the TIMER survey to disentangle the star formation history of the nuclear disc from that of the underlying main disc, which enables us to more accurately determine when the nuclear disc forms. We demonstrate the methodology on the galaxy NGC 1433 -- which we find to host an old bar that is $8.0^{+1.6}_{-1.1}\rm{(sys)}^{+0.2}_{-0.5}\rm{(stat)}$ Gyr old -- and describe a number of tests carried out on both the observational data and numerical simulations. In addition, we present evidence that the nuclear disc of NGC 1433 grows in accordance with an inside-out formation scenario. This methodology is applicable to high-resolution IFS data of barred galaxies with nuclear discs, making it ideally suited for the TIMER survey sample. In the future we will thus be able to determine the bar age for a large sample of galaxies, shedding light on the epoch of disc settling and bar formation., Comment: Accepted for Publication by A&A, 18 pages, 14 figures