Search

Reinforcement learning (RL) is a powerful technique for training intelligent agents, but understanding why these agents make specific decisions can be quite challenging. This lack of transparency in RL models has been a long-standing problem, making it difficult for users to grasp the reasons behind an agent's behaviour. Various approaches have been explored to address this problem, with one promising avenue being reward decomposition (RD). RD is appealing as it sidesteps some of the concerns associated with other methods that attempt to rationalize an agent's behaviour in a post-hoc manner. RD works by exposing various facets of the rewards that contribute to the agent's objectives during training. However, RD alone has limitations as it primarily offers insights based on sub-rewards and does not delve into the intricate cause-and-effect relationships that occur within an RL agent's neural model. In this paper, we present an extension of RD that goes beyond sub-rewards to provide more informative explanations. Our approach is centred on a causal learning framework that leverages information-theoretic measures for explanation objectives that encourage three crucial properties of causal factors: causal sufficiency, sparseness, and orthogonality. These properties help us distill the cause-and-effect relationships between the agent's states and actions or rewards, allowing for a deeper understanding of its decision-making processes. Our framework is designed to generate local explanations and can be applied to a wide range of RL tasks with multiple reward channels. Through a series of experiments, we demonstrate that our approach offers more meaningful and insightful explanations for the agent's action selections., Comment: https://lukaswill.github.io/; Accepted as oral by CLeaR 2024

Explaining the behaviour of intelligent agents learned by reinforcement learning (RL) to humans is challenging yet crucial due to their incomprehensible proprioceptive states, variational intermediate goals, and resultant unpredictability. Moreover, one-step explanations for RL agents can be ambiguous as they fail to account for the agent's future behaviour at each transition, adding to the complexity of explaining robot actions. By leveraging abstracted actions that map to task-specific primitives, we avoid explanations on the movement level. To further improve the transparency and explainability of robotic systems, we propose an explainable Q-Map learning framework that combines reward decomposition (RD) with abstracted action spaces, allowing for non-ambiguous and high-level explanations based on object properties in the task. We demonstrate the effectiveness of our framework through quantitative and qualitative analysis of two robotic scenarios, showcasing visual and textual explanations, from output artefacts of RD explanations, that are easy for humans to comprehend. Additionally, we demonstrate the versatility of integrating these artefacts with large language models (LLMs) for reasoning and interactive querying., Comment: https://lukaswill.github.io

Neural fields are neural networks which map coordinates to a desired signal. When a neural field should jointly model multiple signals, and not memorize only one, it needs to be conditioned on a latent code which describes the signal at hand. Despite being an important aspect, there has been little research on conditioning strategies for neural fields. In this work, we explore the use of neural fields as decoders for 2D semantic segmentation. For this task, we compare three conditioning methods, simple concatenation of the latent code, Feature Wise Linear Modulation (FiLM), and Cross-Attention, in conjunction with latent codes which either describe the full image or only a local region of the image. Our results show a considerable difference in performance between the examined conditioning strategies. Furthermore, we show that conditioning via Cross-Attention achieves the best results and is competitive with a CNN-based decoder for semantic segmentation., Comment: 13 pages, 4 figures, submitted to ICANN 2023

The first population of X-ray binaries (XRBs) is expected to affect the thermal and ionization states of the gas in the early Universe. Although these X-ray sources are predicted to have important implications for high-redshift observable signals, such as the hydrogen 21-cm signal from cosmic dawn and the cosmic X-ray background, their properties are poorly explored, leaving theoretical models largely uninformed. In this paper we model a population of X-ray binaries arising from zero metallicity stars. We explore how their properties depend on the adopted initial mass function (IMF) of primordial stars, finding a strong effect on their number and X-ray production efficiency. We also present scaling relations between XRBs and their X-ray emission with the local star formation rate, which can be used in sub-grid models in numerical simulations to improve the X-ray feedback prescriptions. Specifically, we find that the uniformity and strength of the X-ray feedback in the intergalactic medium is strongly dependant on the IMF. Bottom-heavy IMFs result in a smoother distribution of XRBs, but have a luminosity orders of magnitude lower than more top-heavy IMFs. Top-heavy IMFs lead to more spatially uneven, albeit strong, X-ray emission. An intermediate IMF has a strong X-ray feedback while sustaining an even emission across the intergalactic medium. These differences in X-ray feedback could be probed in the future with measurements of the cosmic dawn 21-cm line of neutral hydrogen, which offers us a new way of constraining population III IMF., Comment: Accepted for publication in MNRAS, 17 pages, 9 figures

The ultra-faint dwarf galaxy Reticulum~II was enriched by a single rare and prolific r-process event. The r-process content of Reticulum~II thus provides a unique opportunity to study metal mixing in a relic first galaxy. Using multi-object high-resolution spectroscopy with VLT/GIRAFFE and Magellan/M2FS, we identify 32 clear spectroscopic member stars and measure abundances of Mg, Ca, Fe, and Ba where possible. We find $72^{+10}_{-12}$% of the stars are r-process-enhanced, with a mean $\left\langle\mbox{[Ba/H]}\right\rangle=-1.68~\pm~0.07$ and unresolved intrinsic dispersion $\sigma_{\rm [Ba/H]} < 0.20$. The homogeneous r-process abundances imply that Ret~II's metals are well-mixed by the time the r-enhanced stars form, which simulations have shown requires at least 100 Myr of metal mixing in between bursts of star formation to homogenize. This is the first direct evidence of bursty star formation in an ultra-faint dwarf galaxy. The homogeneous dilution prefers a prompt and high-yield r-process site, such as collapsar disk winds or prompt neutron star mergers. We also find evidence from [Ba/H] and [Mg/Ca] that the r-enhanced stars in Ret~II formed in the absence of substantial pristine gas accretion, perhaps indicating that ${\approx}70$% of Ret~II stars formed after reionization., Comment: 37 pages, 12 figures, 10 tables, accepted to AJ

Galaxies are thought to reside inside of large gravitationally bound structures of dark matter, so-called haloes. While the smallest of these haloes host no or only a few stars, the biggest host entire clusters of galaxies. Over cosmic history, haloes often collided and merged, forming bigger and bigger structures. Merger trees, i.e., catalogues of haloes evolving and connections between them as they grow and merge, have become a vital tool in describing and understanding the history of cosmological objects such as our Galaxy. Semi-analytical models, built on top of such merger trees, are a common approach for theoretical studies in cosmology. The semi-analytical nature of such models is especially beneficial when the dynamic range in spatial and time scales that need to be considered becomes too large for numerical simulations. Ancient Stars and Local Observables by Tracing Halos (A-SLOTH) is such a semi-analytical model and it is designed to simulate star formation in the early Universe in a fast and accessible way. It uses merger trees, either from numerical simulations or generated by statistical algorithms to describe the history of galaxies. The processes of baryonic physics, in particular gas cooling, star formation and stellar feedback are described with approximations and statistical models. The range of applications for this model is extensive and we, therefore, make it available to the scientific community., Comment: Published in JOSS. Git repo is available at https://gitlab.com/thartwig/asloth

Physically realistic models of stellar spectra are needed in a variety of astronomical studies, from the analysis of fundamental stellar parameters, to studies of exoplanets and stellar populations in galaxies. Here we present a new version of the widely-used radiative transfer code Turbospectrum, which we update with the capacity to perform spectrum synthesis for lines of multiple chemical elements in Non-Local Thermodynamic Equilibrium (NLTE). We use the code in the analysis of metallicites and abundances of the Gaia FGK benchmark stars, using one-dimensional MARCS atmospheric models and the averages of 3D radiation-hydrodynamics simulations of stellar surface convection. We show that the new more physically realistic models offer a better description of the observed data and make the program and the associated microphysics data publicly available, including grids of NLTE departure coefficients for H, O, Na, Mg, Si, Ca, Ti, Mn, Fe, Co, Ni, Sr, and Ba.

The semi-analytical model A-SLOTH (Ancient Stars and Local Observables by Tracing Halos) is the first public code that connects the formation of the first stars and galaxies to observables. After several successful projects with this model, we publish the source code and describe the public version in this paper. The model is based on dark matter merger trees that can either be generated based on Extended Press-Schechter theory or that can be imported from dark matter simulations. On top of these merger trees, A-SLOTH applies analytical recipes for baryonic physics to model the formation of both metal-free and metal-poor stars and the transition between them with unprecedented precision and fidelity. A-SLOTH samples individual stars and includes radiative, chemical, and mechanical feedback. It is calibrated based on six observables, such as the optical depth to Thomson scattering, the stellar mass of the Milky Way and its satellite galaxies, the number of extremely-metal poor stars, and the cosmic star formation rate density at high redshift. A-SLOTH has versatile applications with moderate computational requirements. It can be used to constrain the properties of the first stars and high-z galaxies based on local observables, predicts properties of the oldest and most metal-poor stars in the Milky Way, can serve as a subgrid model for larger cosmological simulations, and predicts next-generation observables of the early Universe, such as supernova rates or gravitational wave events., Comment: Published in ApJ, 35 pages, video summary can be watched at https://www.youtube.com/watch?v=M_JRT2AsOqg and git repository is available at https://gitlab.com/thartwig/asloth

The chemical composition of the Sun is requested in the context of various studies in astrophysics, among them in the calculation of the standard solar models (SSMs), which describe the evolution of the Sun from the pre-main-sequence to its present age. In this work, we provide a critical re-analysis of the solar chemical abundances and corresponding SSMs. For the photospheric values, we employ new high-quality solar observational data collected with the IAG facility, state-of-the art non-equilibrium modelling, new oscillator strengths, and different atmospheric models, including the MARCS model, but also averages based on Stagger and CO5BOLD 3D radiation-hydrodynamics simulations of stellar convection. We perform new calculations of oscillator strengths for transitions in O I and N I. For O I - the critical element for the interior models - calculations are carried out using several independent methods. We find unprecedented agreement between the new estimates of transition probabilities, thus supporting our revised solar oxygen abundance. We also provide new estimates of the noble gas Ne abundance. We investigate our results in comparison with the previous estimates. We discuss the consistency of our photospheric measurements with meteoritic values taking into account systematic and correlated errors. Finally, we provide revised chemical abundances, leading to a new value of the solar photospheric present-day metallicity $Z/X = 0.0225$, and employ them in the calculations of the SSM. We find that the puzzling mismatch between the helioseismic constraints on the solar interior structure and the model is resolved with the new chemical composition., Comment: accepted for publication in A&A

Properties of the first generation of stars (Pop III), such as their initial mass function (IMF), are poorly constrained by observations and have yet to converge between simulations. The cosmological 21-cm signal of neutral hydrogen is predicted to be sensitive to Lyman-band photons produced by these stars, thus providing a unique way to probe the first stellar population. In this paper, we investigate the impacts of the Pop III IMF on the cosmic dawn 21-cm signal via the Wouthuysen-Field effect, Lyman-Werner feedback, Ly-alpha heating, and CMB heating. We calculate the emission spectra of star-forming halos for different IMFs by integrating over individual metal-free stellar spectra, computed from a set of stellar evolution histories and stellar atmospheres, and taking into account variability of the spectra with stellar age. Through this study, we therefore relax two common assumptions: that the zero-age main sequence emission rate of a Pop III star is representative of its lifetime mean emission rate, and that Pop III emission can be treated as instantaneous. Exploring a bottom-heavy, a top-heavy, and intermediate IMFs, we show that variations in the 21-cm signal are driven by stars lighter than 20 solar masses. For the explored models we find maximum relative differences of 59% in the cosmic dawn global 21-cm signal, and 131% between power spectra. Although this impact is modest, precise modelling of the first stars and their evolution is necessary for accurate prediction and interpretation of the 21-cm signal., Comment: 20 pages, 16 figures. Accepted for publication in MNRAS, updated to accepted version

We study the stellar mass-to-halo mass relation at $z=0$ in 30 Milky Way-like systems down to the ultra-faint ($M_* < 10^5 M_\odot$) regime using the semi-analytic model A-SLOTH. A new model allows us to follow star formation and the stochastic stellar feedback from individually sampled Pop II stars. Our fiducial model produces consistent results with the stellar mass-to-halo mass relation derived from abundance matching and the observed cumulative stellar mass function above the observational completeness. We find a plateau in the stellar mass-to-halo mass relation in the ultra-faint regime. The stellar mass of this plateau tells us how many stars formed before supernovae occur and regulate further star formation, which is determined by the Pop~II star formation efficiency. We also find that the number of luminous satellites increases rapidly as $M_*$ decreases until $M_* \approx 10^4 M_\odot$. Finally, we find that the relative streaming velocity between baryons and dark matter at high redshift is important in determining the number of ultra-faint dwarf galaxies at $z=0$. The new model in A-SLOTH provides a framework to study the stellar properties and the formation history of metal-poor stars in Milky Way and its satellites., Comment: 18 pages, 11 figures, 3 tables

We present detailed abundance measurements for 45 globular clusters (GCs) in galaxies in (and, in one case, beyond) the Local Group. The measurements are based on new high-resolution integrated-light spectra of GCs in NGC 185, NGC 205, M31, M33, and NGC 2403, combined with reanalysis of previous observations of GCs in the Fornax dSph, WLM, NGC 147, NGC 6822, and the Milky Way. The GCs cover the range -2.8 < [Fe/H] < -0.1 and we determined abundances for Fe, Na, Mg, Si, Ca, Sc, Ti, Cr, Mn, Ni, Cu, Zn, Zr, Ba, and Eu. Corrections for non local thermodynamic equilibrium effects are included for Na, Mg, Ca, Ti, Mn, Fe, Ni, and Ba. For several of the galaxies, our measurements provide the first quantitative constraints on the detailed composition of their metal-poor stellar populations. Overall, the GCs in different galaxies exhibit remarkably uniform abundance patterns of the alpha-, iron-peak, and neutron-capture elements, with a dispersion of less than 0.1 dex in [alpha/Fe] for the full sample. There is a hint that GCs in dwarf galaxies are slightly less alpha-enhanced (by about 0.04 dex on average) than those in larger galaxies. One GC in M33 (HM33-B) resembles the most metal-rich GCs in the Fornax dSph (Fornax 4) and NGC 6822 (SC7) by having alpha-element abundances closer to scaled-solar values, possibly hinting at an accretion origin. We find that the alpha-element abundances strongly correlate with those of Na, Sc, Ni, and Zn. Several GCs with [Fe/H]<-1.5 are deficient in Mg compared to other alpha-elements. We find no GCs with strongly enhanced r-process abundances as reported for metal-poor stars in some ultra-faint dwarfs and the Magellanic Clouds. The similarity of the abundance patterns for metal-poor GCs in different environments points to similar early enrichment histories and only allow for minor variations in the initial mass function., Comment: 34 pages + 6 appendices. Accepted for publication in Astronomy & Astrophysics

Mapping Cosmic Dawn with 21-cm tomography offers an exciting new window into the era of primordial star formation. However, self-consistent implementation of both the process of star formation and the related 21-cm signal is challenging, due to the multi-scale nature of the problem. In this study, we develop a flexible semi-analytical model to follow the formation of the first stars and the process of gradual transition from primordial to metal-enriched star formation. For this transition we use different in scenarios with varying time-delays (or recovery times) between the first supernovae and the formation of the second generation of stars. We use recovery times between 10 and 100\,Myr and find that these delays have a strong impact on the redshift at which the transition to metal-enriched star formation occurs. We then explore the effect of this transition on the 21-cm signal and find that the recovery time has a distinctive imprint in the signal. Together with an improved understanding of how this time-delay relates to the properties of Population~III stars, future 21-cm observations can give independent constraints on the earliest epoch of star formation., Comment: 17 pages, 11 Figures, submitted to MNRAS, comments very welcome

The lack of observations of abundance patterns originating in pair-instability supernovae has been a long-standing problem in relation to the first stars. This class of supernovae is expected to have an abundance pattern with a strong odd-even effect, making it substantially different from present-day supernovae. In this study, we use a cosmological radiation hydrodynamics simulation to model such supernovae and the subsequent formation of the second generation of stars. We incorporate streaming velocities for the first time. There are 14 star-forming minihalos in our $1\,\mathrm{Mpc}\,h^{-1}$ box, leading to 14 supernovae occurring before redshift $z=19.5$, where we start reducing the complexity of the simulation. Following the explosions, extremely metal-poor stars form in 10 halos via internal and external enrichment, which makes it the most common outcome. Only one halo does not recollapse during the simulations. This result is at tension with the current (lack of) observations of metal-poor stars with pair instability supernova abundance patterns, suggesting that these very massive stars might be rare even in the early Universe. The results from this simulation also give us insights into what drives different modes of recollapse and what determines the mixing behavior of metals after very energetic supernovae., Comment: 21 pages, 17 figures, accepted for publication in ApJ

Stellar and supernova nucleosynthesis in the first few billion years of the cosmic history have set the scene for early structure formation in the Universe, while little is known about their nature. Making use of stellar physical parameters measured by GALAH Data Release 3 with accurate astrometry from the Gaia EDR3, we have selected $\sim 100$ old main-sequence turn-off stars (ages $\gtrsim 12$ Gyrs) with kinematics compatible with the Milky Way stellar halo population in the Solar neighborhood. Detailed homogeneous elemental abundance estimates by GALAH DR3 are compared with supernova yield models of Pop~III (zero-metal) core-collapse supernovae (CCSNe), normal (non-zero-metal) CCSNe, and Type Ia supernovae (SN Ia) to examine which of the individual yields or their combinations best reproduce the observed elemental abundance patterns for each of the old halo stars ("OHS"). We find that the observed abundances in the OHS with [Fe/H]$>-1.5$ are best explained by contributions from both CCSNe and SN~Ia, where the fraction of SN~Ia among all the metal-enriching SNe is up to 10-20 % for stars with high [Mg/Fe] ratios and up to 20-27 % for stars with low [Mg/Fe] ratios, depending on the assumption about the relative fraction of near-Chandrasekhar-mass SNe Ia progenitors. The results suggest that, in the progenitor systems of the OHS with [Fe/H]$>-1.5$, $\sim$ 50-60% of Fe mass originated from normal CCSNe at the earliest phases of the Milky Way formation. These results provide an insight into the birth environments of the oldest stars in the Galactic halo., Comment: 21 pages, 16 Figures, accepted for publication in MNRAS

In domains where computational resources and labeled data are limited, such as in robotics, deep networks with millions of weights might not be the optimal solution. In this paper, we introduce a connectivity scheme for pyramidal architectures to increase their capacity for learning features. Experiments on facial expression recognition of unseen people demonstrate that our approach is a potential candidate for applications with restricted resources, due to good generalization performance and low computational cost. We show that our approach generalizes as well as convolutional architectures in this task but uses fewer trainable parameters and is more robust for low-resolution faces.

Sentiment analysis has transitioned from classifying the sentiment of an entire sentence to providing the contextual information of what targets exist in a sentence, what sentiment the individual targets have, and what the causal words responsible for that sentiment are. However, this has led to elaborate requirements being placed on the datasets needed to train neural networks on the joint triplet task of determining an entity, its sentiment, and the causal words for that sentiment. Requiring this kind of data for training systems is problematic, as they suffer from stacking subjective annotations and domain over-fitting leading to poor model generalisation when applied in new contexts. These problems are also likely to be compounded as we attempt to jointly determine additional contextual elements in the future. To mitigate these problems, we present a hybrid neural-symbolic method utilising a Dependency Tree-LSTM's compositional sentiment parse structure and complementary symbolic rules to correctly extract target-sentiment-cause triplets from sentences without the need for triplet training data. We show that this method has the potential to perform in line with state-of-the-art approaches while also simplifying the data required and providing a degree of interpretability through the Tree-LSTM., Comment: 12 pages, 1 figure

Effectively recognising and applying emotions to interactions is a highly desirable trait for social robots. Implicitly understanding how subjects experience different kinds of actions and objects in the world is crucial for natural HRI interactions, with the possibility to perform positive actions and avoid negative actions. In this paper, we utilize the NICO robot's appearance and capabilities to give the NICO the ability to model a coherent affective association between a perceived auditory stimulus and a temporally asynchronous emotion expression. This is done by combining evaluations of emotional valence from vision and language. NICO uses this information to make decisions about when to extend conversations in order to accrue more affective information if the representation of the association is not coherent. Our primary contribution is providing a NICO robot with the ability to learn the affective associations between a perceived auditory stimulus and an emotional expression. NICO is able to do this for both individual subjects and specific stimuli, with the aid of an emotion-driven dialogue system that rectifies emotional expression incoherences. The robot is then able to use this information to determine a subject's enjoyment of perceived auditory stimuli in a real HRI scenario.

Social robots able to continually learn facial expressions could progressively improve their emotion recognition capability towards people interacting with them. Semi-supervised learning through ensemble predictions is an efficient strategy to leverage the high exposure of unlabelled facial expressions during human-robot interactions. Traditional ensemble-based systems, however, are composed of several independent classifiers leading to a high degree of redundancy, and unnecessary allocation of computational resources. In this paper, we proposed an ensemble based on convolutional networks where the early layers are strong low-level feature extractors, and their representations shared with an ensemble of convolutional branches. This results in a significant drop in redundancy of low-level features processing. Training in a semi-supervised setting, we show that our approach is able to continually learn facial expressions through ensemble predictions using unlabelled samples from different data distributions.

Explaining the outcome of deep learning decisions based on affect is challenging but necessary if we expect social companion robots to interact with users on an emotional level. In this paper, we present a commonsense approach that utilizes an interpretable hybrid neural-symbolic system to associate extracted targets, noun chunks determined to be associated with the expressed emotion, with affective labels from a natural language expression. We leverage a pre-trained neural network that is well adapted to tree and sub-tree processing, the Dependency Tree-LSTM, to learn the affect labels of dynamic targets, determined through symbolic rules, in natural language. We find that making use of the unique properties of the recursive network provides higher accuracy and interpretability when compared to other unstructured and sequential methods for determining target-affect associations in an aspect-based sentiment analysis task., Comment: 6 pages, 5 figures

Recognizing emotions from text in multimodal architectures has yielded promising results, surpassing video and audio modalities under certain circumstances. However, the method by which multimodal data is collected can be significant for recognizing emotional features in language. In this paper, we address the influence data collection methodology has on two multimodal emotion recognition datasets, the IEMOCAP dataset and the OMG-Emotion Behavior dataset, by analyzing textual dataset compositions and emotion recognition accuracy. Experiments with the full IEMOCAP dataset indicate that the composition negatively influences generalization performance when compared to the OMG-Emotion Behavior dataset. We conclude by discussing the impact this may have on HRI experiments., Comment: 2 pages, 2 figures, presented at IROS 2018 Workshop on Language and Robotics

A long sought after goal using chemical abundance patterns derived from metal-poor stars is to understand the Galactic chemical evolution (GCE) and to pin down the nature of the first stars (Pop III). Here, we use a sample of 14 metal-poor stars observed with the high-resolution spectrograph PEPSI at the LBT to derive abundances of 32 elements (34 including limits). We present well-sampled abundance patterns for all stars obtained using local thermodynamic equilibrium (LTE) radiative transfer codes and 1D hydrostatic model atmospheres. It is currently well known that the assumptions of 1D and LTE may hide several issues, thereby introducing biases in our interpretation as to the nature of the first stars and the GCE. Hence, we use non-LTE (NLTE) and correct the abundances using 3D model atmospheres to present a physically more reliable pattern. In order to infer the nature of the first stars, we compare unevolved, cool stars, enriched by a single event (`mono-enriched'), with a set of yield predictions to pin down the mass and energy of the Pop III progenitor. To date, only few bona fide second generation mono-enriched stars are known. A simple x^2-fit may bias our inferred mass and energy just as much as the simple 1D LTE abundance pattern, and we thus carried out our study with an improved fitting technique considering dilution and mixing. Our sample presents Carbon Enhanced Metal-Poor (CEMP) stars, some of which are promising true second generation (mono-enriched) stars. The unevolved, dwarf BD+09_2190 shows a mono-enriched signature which, combined with kinematical data, indicates that it moves in the outer halo and likely has been accreted onto the Milky Way early on. The Pop III progenitor was likely of 25.5M and 0.6 10^51erg (foe)/19.2M and 1.5foe in LTE/NLTE. Finally, we explore the predominant donor and formation site of the rapid and slow neutron-capture elements. Abridged, Comment: 22 pages, 16 figures, online material. Accepted for publication in A&A

To date no metal-free stars have been identified by direct observations. The most common method of constraining their properties is searching the spectra of the most metal-poor stars for the chemical elements created in the first stars and their supernova. In this approach, modelled supernova yields are compared to the observed abundance patterns in extremely metal-poor stars. The method typically only uses the abundance ratios, i.e., the yields are diluted to the observed level. Following the usual assumption of spherical symmetry we compute a simple lower limit of the mass a supernova can mix with and find that it is consistent with all published simulations of early chemical enrichment in the interstellar medium. For three different cases, we demonstrate that this dilution limit can change the conclusions from the abundance fitting. There is a large discrepancy between the dilution found in simulations of SN explosions in minihaloes and the dilution assumed in many abundance fits. Limiting the dilution can significantly alter the likelihood of which supernovae are possible progenitors of observed CEMP-no stars. In particular, some of the faint, very low-yield SNe, which have been suggested as models for the abundance pattern of SMSS0313-6708, cannot explain the measured metal abundances, as their predicted metal yields are too small by two orders of magnitude. Altogether, the new dilution model presented here emphasizes the need to better understand the mixing and dilution behaviour of aspherical SNe., Comment: 10 pages, 6 figures, submitted to MNRAS

The first supernovae enrich the previously pristine gas with metals, out of which the next generation of stars form. Based on hydrodynamical simulations, we develop a new stochastic model to predict the metallicity of star-forming gas in the first galaxies. On average, in internally enriched galaxies, the metals are well mixed with the pristine gas. However, in externally enriched galaxies, the metals can not easily penetrate into the dense gas, which yields a significant metallicity difference between the star-forming and average gas inside a halo. To study the consequences of this effect, we apply a semi-analytical model to Milky Way-like dark matter merger trees and follow stellar fossils from high redshift until the present day with a novel realistic metal mixing recipe. We calibrate the model to reproduce the metallicity distribution function (MDF) at low metallicities and find that a primordial IMF with slope of $\mathrm{d}N/\mathrm{d}M \propto M^{-0.5}$ from $2 Msun$ to $180 Msun$ best reproduces the MDF. Our improved model for inhomogeneous mixing can have a large impact for individual minihalos, but does not significantly influence the modelled MDF at [Fe/H]$\gtrsim -4$ or the best-fitting Pop~III IMF., Comment: 17 pages, 10 figures, resubmitted to ApJ

Ensemble methods, traditionally built with independently trained de-correlated models, have proven to be efficient methods for reducing the remaining residual generalization error, which results in robust and accurate methods for real-world applications. In the context of deep learning, however, training an ensemble of deep networks is costly and generates high redundancy which is inefficient. In this paper, we present experiments on Ensembles with Shared Representations (ESRs) based on convolutional networks to demonstrate, quantitatively and qualitatively, their data processing efficiency and scalability to large-scale datasets of facial expressions. We show that redundancy and computational load can be dramatically reduced by varying the branching level of the ESR without loss of diversity and generalization power, which are both important for ensemble performance. Experiments on large-scale datasets suggest that ESRs reduce the remaining residual generalization error on the AffectNet and FER+ datasets, reach human-level performance, and outperform state-of-the-art methods on facial expression recognition in the wild using emotion and affect concepts., Comment: Accepted at the Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI-20), 1-1, New York, USA

The recognition of emotion and dialogue acts enriches conversational analysis and help to build natural dialogue systems. Emotion interpretation makes us understand feelings and dialogue acts reflect the intentions and performative functions in the utterances. However, most of the textual and multi-modal conversational emotion corpora contain only emotion labels but not dialogue acts. To address this problem, we propose to use a pool of various recurrent neural models trained on a dialogue act corpus, with and without context. These neural models annotate the emotion corpora with dialogue act labels, and an ensemble annotator extracts the final dialogue act label. We annotated two accessible multi-modal emotion corpora: IEMOCAP and MELD. We analyzed the co-occurrence of emotion and dialogue act labels and discovered specific relations. For example, Accept/Agree dialogue acts often occur with the Joy emotion, Apology with Sadness, and Thanking with Joy. We make the Emotional Dialogue Acts (EDA) corpus publicly available to the research community for further study and analysis., Comment: Proceeding of the LREC 2020