@article{13359, abstract = {Dissipative self-assembly is ubiquitous in nature, where it gives rise to complex structures and functions such as self-healing, homeostasis, and camouflage. These phenomena are enabled by the continuous conversion of energy stored in chemical fuels, such as ATP. Over the past decade, an increasing number of synthetic chemically driven systems have been reported that mimic the features of their natural counterparts. At the same time, it has been shown that dissipative self-assembly can also be fueled by light; these optically fueled systems have been developed in parallel to the chemically fueled ones. In this perspective, we critically compare these two classes of systems. Despite the complementarity and fundamental differences between these two modes of dissipative self-assembly, our analysis reveals that multiple analogies exist between chemically and light-fueled systems. We hope that these considerations will facilitate further development of the field of dissipative self-assembly.}, author = {Weißenfels, Maren and Gemen, Julius and Klajn, Rafal}, issn = {2451-9294}, journal = {Chem}, keywords = {Materials Chemistry, Biochemistry (medical), General Chemical Engineering, Environmental Chemistry, Biochemistry, General Chemistry}, number = {1}, pages = {23--37}, publisher = {Elsevier}, title = {{Dissipative self-assembly: Fueling with chemicals versus light}}, doi = {10.1016/j.chempr.2020.11.025}, volume = {7}, year = {2021}, } @inbook{13360, abstract = {Inorganic nanoparticles (NPs) exhibit a wide range of fascinating physicochemical properties, many of which can be controlled by modulating the NP–NP coupling. Controlling the self-assembly of NPs using light has traditionally been achieved by functionalizing their surfaces with monolayers of photoswitchable molecules, which can be reversibly isomerized between two or more states upon exposure to different wavelengths of light. NPs whose assembly can be controlled by light in a reversible fashion can find interesting applications. The chapter deals with systems comprising mixtures of non-photoswitchable NPs and small-molecule photoacids and photobases. Examples of light-controlled self-assembly of NPs hitherto reported have been categorized into six distinct approaches. These are: functionalizing NPs with monolayers of photoswitchable molecules, light-controlled adsorption/desorption of photoswitchable molecules onto NPs, and light-induced electron transfer between the particle's inorganic core and the NP-bound ligands.}, author = {Bian, Tong and Chu, Zonglin and Klajn, Rafal}, booktitle = {Out‐of‐Equilibrium (Supra)molecular Systems and Materials}, editor = {Giuseppone, Nicolas and Walther, Andreas}, isbn = {9783527346158}, pages = {241--273}, publisher = {Wiley}, title = {{Controlling Self‐Assembly of Nanoparticles Using Light}}, doi = {10.1002/9783527821990.ch9}, year = {2021}, } @article{13453, abstract = {Most massive stars are born in binaries close enough for mass transfer episodes. These modify the appearance, structure, and future evolution of both stars. We compute the evolution of a 100-day-period binary, consisting initially of a 25 M⊙ star and a 17 M⊙ star, which experiences stable mass transfer. We focus on the impact of mass accretion on the surface composition, internal rotation, and structure of the accretor. To anchor our models, we show that our accretor broadly reproduces the properties of ζ Ophiuchi, which has long been proposed to have accreted mass before being ejected as a runaway star when the companion exploded. We compare our accretor to models of single rotating stars and find that the later and stronger spin-up provided by mass accretion produces significant differences. Specifically, the core of the accretor retains higher spin at the end of the main sequence, and a convective layer develops that changes its density profile. Moreover, the surface of the accretor star is polluted by CNO-processed material donated by the companion. Our models show effects of mass accretion in binaries that are not captured in single rotating stellar models. This possibly impacts the further evolution (either in a binary or as single stars), the final collapse, and the resulting spin of the compact object.}, author = {Renzo, M. and Götberg, Ylva Louise Linsdotter}, issn = {1538-4357}, journal = {The Astrophysical Journal}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {2}, publisher = {American Astronomical Society}, title = {{Evolution of accretor stars in massive binaries: Broader implications from modeling ζ Ophiuchi}}, doi = {10.3847/1538-4357/ac29c5}, volume = {923}, year = {2021}, } @article{13454, abstract = {Helium star–carbon-oxygen white dwarf (CO WD) binaries are potential single-degenerate progenitor systems of thermonuclear supernovae. Revisiting a set of binary evolution calculations using the stellar evolution code MESA, we refine our previous predictions about which systems can lead to a thermonuclear supernova and then characterize the properties of the helium star donor at the time of explosion. We convert these model properties to near-UV/optical magnitudes assuming a blackbody spectrum and support this approach using a matched stellar atmosphere model. These models will be valuable to compare with pre-explosion imaging for future supernovae, though we emphasize the observational difficulty of detecting extremely blue companions. The pre-explosion source detected in association with SN 2012Z has been interpreted as a helium star binary containing an initially ultra-massive WD in a multiday orbit. However, extending our binary models to initial CO WD masses of up to 1.2 M⊙, we find that these systems undergo off-center carbon ignitions and thus are not expected to produce thermonuclear supernovae. This tension suggests that, if SN 2012Z is associated with a helium star–WD binary, then the pre-explosion optical light from the system must be significantly modified by the binary environment and/or the WD does not have a carbon-rich interior composition.}, author = {Wong, Tin Long Sunny and Schwab, Josiah and Götberg, Ylva Louise Linsdotter}, issn = {1538-4357}, journal = {The Astrophysical Journal}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {2}, publisher = {American Astronomical Society}, title = {{Pre-explosion properties of Helium star donors to thermonuclear supernovae}}, doi = {10.3847/1538-4357/ac27ae}, volume = {922}, year = {2021}, } @article{13455, abstract = {The majority of massive stars live in binary or multiple systems and will interact with a companion during their lifetimes, which helps to explain the observed diversity of core-collapse supernovae. Donor stars in binary systems can lose most of their hydrogen-rich envelopes through mass transfer. As a result, not only are the surface properties affected, but so is the core structure. However, most calculations of the core-collapse properties of massive stars rely on single-star models. We present a systematic study of the difference between the pre-supernova structures of single stars and stars of the same initial mass (11–21 M⊙) that have been stripped due to stable post-main-sequence mass transfer at solar metallicity. We present the pre-supernova core composition with novel diagrams that give an intuitive representation of the isotope distribution. As shown in previous studies, at the edge of the carbon-oxygen core, the binary-stripped star models contain an extended gradient of carbon, oxygen, and neon. This layer remains until core collapse and is more extended in mass for higher initial stellar masses. It originates from the receding of the convective helium core during core helium burning in binary-stripped stars, which does not occur in single-star models. We find that this same evolutionary phase leads to systematic differences in the final density and nuclear energy generation profiles. Binary-stripped star models have systematically higher total masses of carbon at the moment of core collapse compared to single-star models, which likely results in systematically different supernova yields. In about half of our models, the silicon-burning and oxygen-rich layers merge after core silicon burning. We discuss the implications of our findings for the “explodability”, supernova observations, and nucleosynthesis of these stars. Our models are publicly available and can be readily used as input for detailed supernova simulations.}, author = {Laplace, E. and Justham, S. and Renzo, M. and Götberg, Ylva Louise Linsdotter and Farmer, R. and Vartanyan, D. and de Mink, S. E.}, issn = {1432-0746}, journal = {Astronomy & Astrophysics}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, publisher = {EDP Sciences}, title = {{Different to the core: The pre-supernova structures of massive single and binary-stripped stars}}, doi = {10.1051/0004-6361/202140506}, volume = {656}, year = {2021}, } @article{13456, abstract = {While most simulations of the epoch of reionization have focused on single-stellar populations in star-forming dwarf galaxies, products of binary evolution are expected to significantly contribute to emissions of hydrogen-ionizing photons. Among these products are stripped stars (or helium stars), which have their envelopes stripped from interactions with binary companions, leaving an exposed helium core. Previous work has suggested these stripped stars can dominate the Lyman Continuum (LyC) photon output of high-redshift, low-luminosity galaxies post-starburst. Other sources of hard radiation in the early universe include zero-metallicity Population iii stars, which may have similar spectral energy distribution (SED) properties to galaxies with radiation dominated by stripped-star emissions. Here, we use four metrics (the power-law exponent over wavelength intervals 240–500 Å, 600–900 Å, and 1200–2000 Å, and the ratio of total luminosity in FUV wavelengths to LyC wavelengths) to compare the SEDs of simulated galaxies with only single-stellar evolution, galaxies containing stripped stars, and galaxies containing Population iii stars, with four different initial mass functions (IMFs). We find that stripped stars significantly alter SEDs in the LyC range of galaxies at the epoch of reionization. SEDs in galaxies with stripped stars have lower power-law indices in the LyC range and lower FUV to LyC luminosity ratios. These differences in SEDs are present at all considered luminosities (${M}_{\mathrm{UV}}\gt -15$, AB system), and are most pronounced for lower-luminosity galaxies. Intrinsic SEDs as well as those with interstellar medium absorption of galaxies with stripped stars and Population iii stars are found to be distinct for all tested Population iii IMFs.}, author = {Berzin, Elizabeth and Secunda, Amy and Cen, Renyue and Menegas, Alexander and Götberg, Ylva Louise Linsdotter}, issn = {1538-4357}, journal = {The Astrophysical Journal}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {1}, publisher = {American Astronomical Society}, title = {{Spectral signatures of population III and envelope-stripped stars in galaxies at the epoch of reionization}}, doi = {10.3847/1538-4357/ac0af6}, volume = {918}, year = {2021}, } @article{13457, abstract = {Context. Observations of massive stars in open clusters younger than ∼8 Myr have shown that a majority of them are in binary systems, most of which will interact during their life. While these can be used as a proxy of the initial multiplicity properties, studying populations of massive stars older than ∼20 Myr allows us to probe the outcome of these interactions after a significant number of systems have experienced mass and angular momentum transfer and may even have merged. Aims. Using multi-epoch integral-field spectroscopy, we aim to investigate the multiplicity properties of the massive-star population in the dense core of the ∼40 Myr old cluster NGC 330 in the Small Magellanic Cloud in order to search for possible imprints of stellar evolution on the multiplicity properties. Methods. We obtained six epochs of VLT/MUSE observations operated in wide-field mode with the extended wavelength setup and supported by adaptive optics. We extracted spectra and measured radial velocities for stars brighter than mF814W = 19. We identified single-lined spectroscopic binaries through significant RV variability with a peak-to-peak amplitude larger than 20 km s−1. We also identified double-lined spectroscopic binaries, and quantified the observational biases for binary detection. In particular, we took into account that binary systems with similar line strengths are difficult to detect in our data set. Results. The observed spectroscopic binary fraction among stars brighter than mF814W = 19 (approximately 5.5 M⊙ on the main sequence) is fSBobs = 13.2 ± 2.0%. Considering period and mass ratio ranges from log(P) = 0.15−3.5 (about 1.4 to 3160 d), q = 0.1−1.0, and a representative set of orbital parameter distributions, we find a bias-corrected close binary fraction of fcl = 34−7+8%. This fraction seems to decline for the fainter stars, which indicates either that the close binary fraction drops in the B-type domain, or that the period distribution becomes more heavily weighted toward longer orbital periods. We further find that both fractions vary strongly in different regions of the color-magnitude diagram, which corresponds to different evolutionary stages. This probably reveals the imprint of the binary history of different groups of stars. In particular, we find that the observed spectroscopic binary fraction of Be stars (fSBobs = 2 ± 2%) is significantly lower than that of B-type stars (fSBobs = 9 ± 2%). Conclusions. We provide the first homogeneous radial velocity study of a large sample of B-type stars at a low metallicity ([Fe/H] ≲ −1.0). The overall bias-corrected close binary fraction (log(P) < 3.5 d) of the B-star population in NGC 330 is lower than the fraction reported for younger Galactic and Large Magellanic Cloud clusters in previous works. More data are needed, however, to establish whether the observed differences are caused by an age or a metallicity effect.}, author = {Bodensteiner, J. and Sana, H. and Wang, C. and Langer, N. and Mahy, L. and Banyard, G. and de Koter, A. and de Mink, S. E. and Evans, C. J. and Götberg, Ylva Louise Linsdotter and Patrick, L. R. and Schneider, F. R. N. and Tramper, F.}, issn = {1432-0746}, journal = {Astronomy & Astrophysics}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, publisher = {EDP Sciences}, title = {{The young massive SMC cluster NGC 330 seen by MUSE. II. Multiplicity properties of the massive-star population}}, doi = {10.1051/0004-6361/202140507}, volume = {652}, year = {2021}, } @article{13458, abstract = {Most massive stars experience binary interactions in their lifetimes that can alter both the surface and core structure of the stripped star with significant effects on their ultimate fate as core-collapse supernovae. However, core-collapse supernovae simulations to date have focused almost exclusively on the evolution of single stars. We present a systematic simulation study of single and binary-stripped stars with the same initial mass as candidates for core-collapse supernovae (11–21 M⊙). Generally, we find that binary-stripped stars core tend to have a smaller compactness parameter, with a more prominent, deeper silicon/oxygen interface, and explode preferentially to the corresponding single stars of the same initial mass. Such a dichotomy of behavior between these two modes of evolution would have important implications for supernovae statistics, including the final neutron star masses, explosion energies, and nucleosynthetic yields. Binary-stripped remnants are also well poised to populate the possible mass gap between the heaviest neutron stars and the lightest black holes. Our work presents an improvement along two fronts, as we self-consistently account for the pre-collapse stellar evolution and the subsequent explosion outcome. Even so, our results emphasize the need for more detailed stellar evolutionary models to capture the sensitive nature of explosion outcome.}, author = {Vartanyan, David and Laplace, Eva and Renzo, Mathieu and Götberg, Ylva Louise Linsdotter and Burrows, Adam and de Mink, Selma E.}, issn = {2041-8213}, journal = {The Astrophysical Journal Letters}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {1}, publisher = {American Astronomical Society}, title = {{Binary-stripped stars as core-collapse supernovae progenitors}}, doi = {10.3847/2041-8213/ac0b42}, volume = {916}, year = {2021}, } @article{13459, abstract = {The B emission-line stars are rapid rotators that were probably spun up by mass and angular momentum accretion through mass transfer in an interacting binary. Mass transfer will strip the donor star of its envelope to create a small and hot subdwarf remnant. Here we report on Hubble Space Telescope/STIS far-ultraviolet spectroscopy of a sample of Be stars that reveals the presence of the hot sdO companion through the calculation of cross-correlation functions of the observed and model spectra. We clearly detect the spectral signature of the sdO star in 10 of the 13 stars in the sample, and the spectral signals indicate that the sdO stars are hot, relatively faint, and slowly rotating as predicted by models. A comparison of their temperatures and radii with evolutionary tracks indicates that the sdO stars occupy the relatively long-lived, He-core burning stage. Only 1 of the 10 detections was a known binary prior to this investigation, which emphasizes the difficulty of finding such Be+sdO binaries through optical spectroscopy. However, these results and others indicate that many Be stars probably host hot subdwarf companions.}, author = {Wang, Luqian and Gies, Douglas R. and Peters, Geraldine J. and Götberg, Ylva Louise Linsdotter and Chojnowski, S. Drew and Lester, Kathryn V. and Howell, Steve B.}, issn = {1538-3881}, journal = {The Astronomical Journal}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {5}, publisher = {American Astronomical Society}, title = {{The detection and characterization of Be+sdO binaries from HST/STIS FUV spectroscopy}}, doi = {10.3847/1538-3881/abf144}, volume = {161}, year = {2021}, } @article{13995, abstract = {Shape resonances play a central role in many areas of science, but the real-time measurement of the associated many-body dynamics remains challenging. Here, we present measurements of recoil frame angle-resolved photoionization delays in the vicinity of shape resonances of CF4. This technique provides insights into the spatiotemporal photoionization dynamics of molecular shape resonances. We find delays of up to ∼600 as in the ionization out of the highest occupied molecular orbital (HOMO) with a strong dependence on the emission direction and a pronounced asymmetry along the dissociation axis. Comparison with quantum-scattering calculations traces the asymmetries to the interference of a small subset of partial waves at low kinetic energies and, additionally, to the interference of two overlapping shape resonances in the HOMO-1 channel. Our experimental and theoretical results establish a broadly applicable approach to space- and time-resolved photoionization dynamics in the molecular frame.}, author = {Heck, Saijoscha and Baykusheva, Denitsa Rangelova and Han, Meng and Ji, Jia-Bao and Perry, Conaill and Gong, Xiaochun and Wörner, Hans Jakob}, issn = {2375-2548}, journal = {Science Advances}, keywords = {Multidisciplinary}, number = {49}, publisher = {American Association for the Advancement of Science}, title = {{Attosecond interferometry of shape resonances in the recoil frame of CF4}}, doi = {10.1126/sciadv.abj8121}, volume = {7}, year = {2021}, } @article{13996, abstract = {We report the observation of an anomalous nonlinear optical response of the prototypical three-dimensional topological insulator bismuth selenide through the process of high-order harmonic generation. We find that the generation efficiency increases as the laser polarization is changed from linear to elliptical, and it becomes maximum for circular polarization. With the aid of a microscopic theory and a detailed analysis of the measured spectra, we reveal that such anomalous enhancement encodes the characteristic topology of the band structure that originates from the interplay of strong spin–orbit coupling and time-reversal symmetry protection. The implications are in ultrafast probing of topological phase transitions, light-field driven dissipationless electronics, and quantum computation.}, author = {Baykusheva, Denitsa Rangelova and Chacón, Alexis and Lu, Jian and Bailey, Trevor P. and Sobota, Jonathan A. and Soifer, Hadas and Kirchmann, Patrick S. and Rotundu, Costel and Uher, Ctirad and Heinz, Tony F. and Reis, David A. and Ghimire, Shambhu}, issn = {1530-6992}, journal = {Nano Letters}, keywords = {Mechanical Engineering, Condensed Matter Physics, General Materials Science, General Chemistry, Bioengineering}, number = {21}, pages = {8970--8978}, publisher = {American Chemical Society}, title = {{All-optical probe of three-dimensional topological insulators based on high-harmonic generation by circularly polarized laser fields}}, doi = {10.1021/acs.nanolett.1c02145}, volume = {21}, year = {2021}, } @article{13997, abstract = {We investigate theoretically the strong-field regime of light-matter interactions in the topological-insulator class of quantum materials. In particular, we focus on the process of nonperturbative high-order harmonic generation from the paradigmatic three-dimensional topological insulator bismuth selenide (Bi2Se3) subjected to intense midinfrared laser fields. We analyze the contributions from the spin-orbit-coupled bulk states and the topological surface bands separately and reveal a major difference in how their harmonic yields depend on the ellipticity of the laser field. Bulk harmonics show a monotonic decrease in their yield as the ellipticity increases, in a manner reminiscent of high harmonic generation in gaseous media. However, the surface contribution exhibits a highly nontrivial dependence, culminating with a maximum for circularly polarized fields. We attribute the observed anomalous behavior to (i) the enhanced amplitude and the circular pattern of the interband dipole and the Berry connections in the vicinity of the Dirac point and (ii) the influence of the higher-order, hexagonal warping terms in the Hamiltonian, which are responsible for the hexagonal deformation of the energy surface at higher momenta. The latter are associated directly with spin-orbit-coupling parameters. Our results thus establish the sensitivity of strong-field-driven high harmonic emission to the topology of the band structure as well as to the manifestations of spin-orbit interaction.}, author = {Baykusheva, Denitsa Rangelova and Chacón, Alexis and Kim, Dasol and Kim, Dong Eon and Reis, David A. and Ghimire, Shambhu}, issn = {2469-9934}, journal = {Physical Review A}, number = {2}, publisher = {American Physical Society}, title = {{Strong-field physics in three-dimensional topological insulators}}, doi = {10.1103/physreva.103.023101}, volume = {103}, year = {2021}, } @unpublished{14097, abstract = {UVEX is a proposed medium class Explorer mission designed to provide crucial missing capabilities that will address objectives central to a broad range of modern astrophysics. The UVEX design has two co-aligned wide-field imagers operating in the FUV and NUV and a powerful broadband medium resolution spectrometer. In its two-year baseline mission, UVEX will perform a multi-cadence synoptic all-sky survey 50/100 times deeper than GALEX in the NUV/FUV, cadenced surveys of the Large and Small Magellanic Clouds, rapid target of opportunity followup, as well as spectroscopic followup of samples of stars and galaxies. The science program is built around three pillars. First, UVEX will explore the low-mass, low-metallicity galaxy frontier through imaging and spectroscopic surveys that will probe key aspects of the evolution of galaxies by understanding how star formation and stellar evolution at low metallicities affect the growth and evolution of low-metallicity, low-mass galaxies in the local universe. Such galaxies contain half the mass in the local universe, and are analogs for the first galaxies, but observed at distances that make them accessible to detailed study. Second, UVEX will explore the dynamic universe through time-domain surveys and prompt spectroscopic followup capability will probe the environments, energetics, and emission processes in the early aftermaths of gravitational wave-discovered compact object mergers, discover hot, fast UV transients, and diagnose the early stages of stellar explosions. Finally, UVEX will become a key community resource by leaving a large all-sky legacy data set, enabling a wide range of scientific studies and filling a gap in the new generation of wide-field, sensitive optical and infrared surveys provided by the Rubin, Euclid, and Roman observatories. This paper discusses the scientific potential of UVEX, and the broad scientific program.}, author = {Kulkarni, S. R. and Harrison, Fiona A. and Grefenstette, Brian W. and Earnshaw, Hannah P. and Andreoni, Igor and Berg, Danielle A. and Bloom, Joshua S. and Cenko, S. Bradley and Chornock, Ryan and Christiansen, Jessie L. and Coughlin, Michael W. and Criswell, Alexander Wuollet and Darvish, Behnam and Das, Kaustav K. and De, Kishalay and Dessart, Luc and Dixon, Don and Dorsman, Bas and Kareem El-Badry, Kareem El-Badry and Evans, Christopher and Ford, K. E. Saavik and Fremling, Christoffer and Gansicke, Boris T. and Gezari, Suvi and Götberg, Ylva Louise Linsdotter and Green, Gregory M. and Graham, Matthew J. and Heida, Marianne and Ho, Anna Y. Q. and Jaodand, Amruta D. and Christopher M. Johns-Krull, Christopher M. Johns-Krull and Kasliwal, Mansi M. and Lazzarini, Margaret and Lu, Wenbin and Margutti, Raffaella and Martin, D. Christopher and Masters, Daniel Charles and McKernan, Barry and Naze, Yael and Nissanke, Samaya M. and Parazin, B. and Perley, Daniel A. and Phinney, E. Sterl and Piro, Anthony L. and Raaijmakers, G. and Rauw, Gregor and Rodriguez, Antonio C. and Sana, Hugues and Senchyna, Peter and Singer, Leo P. and Spake, Jessica J. and Stassun, Keivan G. and Stern, Daniel and Teplitz, Harry I. and Weisz, Daniel R. and Yao, Yuhan}, booktitle = {arXiv}, title = {{Science with the ultraviolet explorer (UVEX)}}, doi = {10.48550/arXiv.2111.15608}, year = {2021}, } @article{14117, abstract = {The two fields of machine learning and graphical causality arose and are developed separately. However, there is, now, cross-pollination and increasing interest in both fields to benefit from the advances of the other. In this article, we review fundamental concepts of causal inference and relate them to crucial open problems of machine learning, including transfer and generalization, thereby assaying how causality can contribute to modern machine learning research. This also applies in the opposite direction: we note that most work in causality starts from the premise that the causal variables are given. A central problem for AI and causality is, thus, causal representation learning, that is, the discovery of high-level causal variables from low-level observations. Finally, we delineate some implications of causality for machine learning and propose key research areas at the intersection of both communities.}, author = {Scholkopf, Bernhard and Locatello, Francesco and Bauer, Stefan and Ke, Nan Rosemary and Kalchbrenner, Nal and Goyal, Anirudh and Bengio, Yoshua}, issn = {1558-2256}, journal = {Proceedings of the IEEE}, keywords = {Electrical and Electronic Engineering}, number = {5}, pages = {612--634}, publisher = {Institute of Electrical and Electronics Engineers}, title = {{Toward causal representation learning}}, doi = {10.1109/jproc.2021.3058954}, volume = {109}, year = {2021}, } @inproceedings{14176, abstract = {Intensive care units (ICU) are increasingly looking towards machine learning for methods to provide online monitoring of critically ill patients. In machine learning, online monitoring is often formulated as a supervised learning problem. Recently, contrastive learning approaches have demonstrated promising improvements over competitive supervised benchmarks. These methods rely on well-understood data augmentation techniques developed for image data which do not apply to online monitoring. In this work, we overcome this limitation by supplementing time-series data augmentation techniques with a novel contrastive learning objective which we call neighborhood contrastive learning (NCL). Our objective explicitly groups together contiguous time segments from each patient while maintaining state-specific information. Our experiments demonstrate a marked improvement over existing work applying contrastive methods to medical time-series.}, author = {Yèche, Hugo and Dresdner, Gideon and Locatello, Francesco and Hüser, Matthias and Rätsch, Gunnar}, booktitle = {Proceedings of 38th International Conference on Machine Learning}, location = {Virtual}, pages = {11964--11974}, publisher = {ML Research Press}, title = {{Neighborhood contrastive learning applied to online patient monitoring}}, volume = {139}, year = {2021}, } @inproceedings{14177, abstract = {The focus of disentanglement approaches has been on identifying independent factors of variation in data. However, the causal variables underlying real-world observations are often not statistically independent. In this work, we bridge the gap to real-world scenarios by analyzing the behavior of the most prominent disentanglement approaches on correlated data in a large-scale empirical study (including 4260 models). We show and quantify that systematically induced correlations in the dataset are being learned and reflected in the latent representations, which has implications for downstream applications of disentanglement such as fairness. We also demonstrate how to resolve these latent correlations, either using weak supervision during training or by post-hoc correcting a pre-trained model with a small number of labels.}, author = {Träuble, Frederik and Creager, Elliot and Kilbertus, Niki and Locatello, Francesco and Dittadi, Andrea and Goyal, Anirudh and Schölkopf, Bernhard and Bauer, Stefan}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, location = {Virtual}, pages = {10401--10412}, publisher = {ML Research Press}, title = {{On disentangled representations learned from correlated data}}, volume = {139}, year = {2021}, } @inproceedings{14178, abstract = {Learning meaningful representations that disentangle the underlying structure of the data generating process is considered to be of key importance in machine learning. While disentangled representations were found to be useful for diverse tasks such as abstract reasoning and fair classification, their scalability and real-world impact remain questionable. We introduce a new high-resolution dataset with 1M simulated images and over 1,800 annotated real-world images of the same setup. In contrast to previous work, this new dataset exhibits correlations, a complex underlying structure, and allows to evaluate transfer to unseen simulated and real-world settings where the encoder i) remains in distribution or ii) is out of distribution. We propose new architectures in order to scale disentangled representation learning to realistic high-resolution settings and conduct a large-scale empirical study of disentangled representations on this dataset. We observe that disentanglement is a good predictor for out-of-distribution (OOD) task performance.}, author = {Dittadi, Andrea and Träuble, Frederik and Locatello, Francesco and Wüthrich, Manuel and Agrawal, Vaibhav and Winther, Ole and Bauer, Stefan and Schölkopf, Bernhard}, booktitle = {The Ninth International Conference on Learning Representations}, location = {Virtual}, title = {{On the transfer of disentangled representations in realistic settings}}, year = {2021}, } @inproceedings{14179, abstract = {Self-supervised representation learning has shown remarkable success in a number of domains. A common practice is to perform data augmentation via hand-crafted transformations intended to leave the semantics of the data invariant. We seek to understand the empirical success of this approach from a theoretical perspective. We formulate the augmentation process as a latent variable model by postulating a partition of the latent representation into a content component, which is assumed invariant to augmentation, and a style component, which is allowed to change. Unlike prior work on disentanglement and independent component analysis, we allow for both nontrivial statistical and causal dependencies in the latent space. We study the identifiability of the latent representation based on pairs of views of the observations and prove sufficient conditions that allow us to identify the invariant content partition up to an invertible mapping in both generative and discriminative settings. We find numerical simulations with dependent latent variables are consistent with our theory. Lastly, we introduce Causal3DIdent, a dataset of high-dimensional, visually complex images with rich causal dependencies, which we use to study the effect of data augmentations performed in practice.}, author = {Kügelgen, Julius von and Sharma, Yash and Gresele, Luigi and Brendel, Wieland and Schölkopf, Bernhard and Besserve, Michel and Locatello, Francesco}, booktitle = {Advances in Neural Information Processing Systems}, isbn = {9781713845393}, location = {Virtual}, pages = {16451--16467}, title = {{Self-supervised learning with data augmentations provably isolates content from style}}, volume = {34}, year = {2021}, } @inproceedings{14180, abstract = {Modern neural network architectures can leverage large amounts of data to generalize well within the training distribution. However, they are less capable of systematic generalization to data drawn from unseen but related distributions, a feat that is hypothesized to require compositional reasoning and reuse of knowledge. In this work, we present Neural Interpreters, an architecture that factorizes inference in a self-attention network as a system of modules, which we call \emph{functions}. Inputs to the model are routed through a sequence of functions in a way that is end-to-end learned. The proposed architecture can flexibly compose computation along width and depth, and lends itself well to capacity extension after training. To demonstrate the versatility of Neural Interpreters, we evaluate it in two distinct settings: image classification and visual abstract reasoning on Raven Progressive Matrices. In the former, we show that Neural Interpreters perform on par with the vision transformer using fewer parameters, while being transferrable to a new task in a sample efficient manner. In the latter, we find that Neural Interpreters are competitive with respect to the state-of-the-art in terms of systematic generalization. }, author = {Rahaman, Nasim and Gondal, Muhammad Waleed and Joshi, Shruti and Gehler, Peter and Bengio, Yoshua and Locatello, Francesco and Schölkopf, Bernhard}, booktitle = {Advances in Neural Information Processing Systems}, isbn = {9781713845393}, location = {Virtual}, pages = {10985--10998}, title = {{Dynamic inference with neural interpreters}}, volume = {34}, year = {2021}, } @inproceedings{14181, abstract = {Variational Inference makes a trade-off between the capacity of the variational family and the tractability of finding an approximate posterior distribution. Instead, Boosting Variational Inference allows practitioners to obtain increasingly good posterior approximations by spending more compute. The main obstacle to widespread adoption of Boosting Variational Inference is the amount of resources necessary to improve over a strong Variational Inference baseline. In our work, we trace this limitation back to the global curvature of the KL-divergence. We characterize how the global curvature impacts time and memory consumption, address the problem with the notion of local curvature, and provide a novel approximate backtracking algorithm for estimating local curvature. We give new theoretical convergence rates for our algorithms and provide experimental validation on synthetic and real-world datasets.}, author = {Dresdner, Gideon and Shekhar, Saurav and Pedregosa, Fabian and Locatello, Francesco and Rätsch, Gunnar}, booktitle = {Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence}, location = {Montreal, Canada}, pages = {2337--2343}, publisher = {International Joint Conferences on Artificial Intelligence}, title = {{Boosting variational inference with locally adaptive step-sizes}}, doi = {10.24963/ijcai.2021/322}, year = {2021}, }