@inproceedings{18061, abstract = {Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing, bringing faster and more accurate models, at the cost of massive parameter counts. For example, the SwitchTransformer-c2048 model has 1.6 trillion parameters, requiring 3.2TB of accelerator memory to run efficiently, which makes practical deployment challenging and expensive. In this paper, we present a solution to this memory problem, in form of a new compression and execution framework called QMoE. Specifically, QMoE consists of a scalable algorithm which accurately compresses trillion-parameter MoEs to less than 1 bit per parameter, in a custom format co-designed with bespoke GPU decoding kernels to facilitate efficient end-to-end compressed inference, with minor runtime overheads relative to uncompressed execution. Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB (20x compression, 0.8 bits per parameter) at only minor accuracy loss, in less than a day on a single GPU. This enables, for the first time, the execution of a trillion-parameter model on affordable commodity hardware, like a single server with 4x NVIDIA A6000 or 8x NVIDIA 3090 GPUs, at less than 5% runtime overhead relative to ideal uncompressed inference. The anonymized code is available at: github.com/mlsys24-qmoe/qmoe.}, author = {Frantar, Elias and Alistarh, Dan-Adrian}, booktitle = {Proceedings of Machine Learning and Systems}, location = {Santa Clara, CA, United States}, title = {{QMoE: Sub-1-bit compression of trillion parameter models}}, volume = {6}, year = {2024}, } @phdthesis{17490, abstract = {Deep learning is essential in numerous applications nowadays, with many recent advancements made possible by training very large models. Despite their broad applicability, training neural networks is often time-intensive, and it is usually impractical to manage large models and datasets on a single machine. To address these issues, distributed deep learning training has become increasingly important. However, distributed training requires synchronization among nodes, and the mini-batch stochastic gradient descent algorithm places a significant load on network connections. A possible solution to tackle the synchronization bottleneck is to reduce a message size by lossy compression. In this thesis, we investigate systems and algorithmic approaches to communication compression during training. From the systems perspective, we demonstrate that a common approach of expensive hardware overprovisioning can be replaced through a thorough system design. We introduce a framework that introduces efficient software support for compressed communication in machine learning applications, applicable to both multi-GPU single-node training and larger-scale multi-node training. Our framework integrates with popular ML frameworks, providing up to 3x speedups for multi-GPU nodes based on commodity hardware and order-of-magnitude improvements in the multi-node setting, with negligible impact on accuracy. Also, we consider an application of our framework to different communication schemes, such as Fully Sharded Data Parallel. We provide strong convergence guarantees for the compression in such a setup. Empirical validation shows that our method preserves model accuracy for GPT-family models with up to 1.3 billion parameters, while completely removing the communication bottlenecks of non-compressed alternatives, providing up to 2.2x speedups end-to-end. From the algorithmic side, we propose a general framework that dynamically adjusts the degree of compression across a model's layers during training. This approach enhances overall compression and results in significant speedups without compromising accuracy. Our algorithm utilizes an adaptive algorithm that automatically selects the optimal compression parameters for model layers, ensuring the best compression ratio while adhering to an error constraint. Our method is effective across all existing families of compression methods. It achieves up to 2.5x faster training and up to a 5x improvement in compression compared to efficient implementations of current approaches. Additionally, LGreCo can complement existing adaptive algorithms. }, author = {Markov, Ilia}, issn = {2663-337X}, pages = {102}, publisher = {Institute of Science and Technology Austria}, title = {{Communication-efficient distributed training of deep neural networks : An algorithms and systems perspective}}, doi = {10.15479/at:ista:17490}, year = {2024}, } @inproceedings{17456, abstract = {Data-parallel distributed training of deep neural networks (DNN) has gained very widespread adoption, but can still experience communication bottlenecks. To address this issue, entire families of compression mechanisms have been developed, including quantization, sparsification, and low-rank approximation, some of which are seeing significant practical adoption. Despite this progress, almost all known compression schemes apply compression uniformly across DNN layers, although layers are heterogeneous in terms of parameter count and their impact on model accuracy.In this work, we provide a general framework for adapting the degree of compression across the model's layers dynamically during training, improving the overall compression, while leading to substantial speedups, without sacrificing accuracy. Our framework, called L-GreCo, is based on an adaptive algorithm, which automatically picks the optimal compression parameters for model layers guaranteeing the best compression ratio while satisfying an error constraint. Extensive experiments over image classification and language modeling tasks shows that L-GreCo is effective across all existing families of compression methods, and achieves up to 2.5 × training speedup and up to 5 × compression improvement over efficient implementations of existing approaches, while recovering full accuracy. Moreover, L-GreCo is complementary to existing adaptive algorithms, improving their compression ratio by 50\% and practical throughput by 66\%. An anonymized implementation is available at https://github.com/LGrCo/L-GreCo.}, author = {Markov, Ilia and Alimohammadi, Kaveh and Frantar, Elias and Alistarh, Dan-Adrian}, booktitle = {Proceedings of Machine Learning and Systems }, editor = {Gibbons, P. and Pekhimenko, G. and De Sa, C.}, location = {Athens, Greece}, publisher = {Association for Computing Machinery}, title = {{L-GreCo: Layerwise-adaptive gradient compression for efficient data-parallel deep learning}}, volume = {6}, year = {2024}, } @phdthesis{18766, abstract = {Poxviruses are large pleomorphic double-stranded DNA viruses that include well known members such as variola virus, the causative agent of smallpox, Mpox virus, as well as Vaccinia virus (VACV), which serves as a vaccination strain for formerly mentioned viruses. VACV is a valuable model for studying large pleomorphic DNA viruses in general and poxviruses specifically, as many features, such as core morphology and structural proteins, are well conserved within this family. Despite decades of research, our understanding of the structural components and proteins that comprise the poxvirus core in mature virions remains limited. Although major core proteins were identified via indirect experimental evidence, the core's complexity, with its large size, structure and number of involved proteins, has hindered efforts to achieve high-resolution insights and to define the roles of the individual proteins. The specific protein composition of the core's individual layers, including the palisade layer and the inner core wall, has remained unclear. In this study, we have merged multiple approaches, including single particle cryo electron microscopy of purified virus cores, cryo-electron tomography and subtomogram averaging of mature virions and molecular modeling to elucidate the structural determinants of the VACV core. Due to the lack of experimentally derived structures, either in situ or reconstituted in vitro, we used Alphafold to predict models of the putative major core protein candidates, A10, 23k, A3, A4, and L4. Our results show that the VACV core is composed of several layers with varying local symmetries, forming more intricate interactions than observed previously. This allowed us to identify several molecular building blocks forming the viral core lattice. In particular, we identified trimers of protein A10 as a major core structure that forms the palisade layer of the viral core. Additionally, we revealed that six petals of a flower shaped core pore within the core wall are composed of A10 trimers. Furthermore, we obtained a cryo-EM density for the inner core wall that could potentially accommodate an A3 dimer. Integrating descriptions of protein interactions from previous studies enabled us to provide a detailed structural model of the poxvirus core wall, and our findings indicate that the interactions within A10 trimers are likely consistent across orthopox- and parapoxviruses. This combined application of cryo-SPA and cryo-ET can help overcome obstacles in studying complex virus structures in the future, including their key assembly proteins, interactions, and the formation into a core lattice. Our work provides important fundamental new insights into poxvirus core architecture, also considering the recent re-emergence of poxviruses.}, author = {Datler, Julia}, isbn = {978-3-99078-049-7}, issn = {2663-337X}, keywords = {cryo-EM, cryo-ET, cryo-SPA, Structural Virology, Poxvirus, Vaccinia Virus, Structural Biology}, pages = {106}, publisher = {Institute of Science and Technology Austria}, title = {{Elucidating the structural determinants of the poxvirus core using multi-modal cryo-EM}}, doi = {10.15479/at:ista:18766}, year = {2024}, } @article{14979, abstract = {Poxviruses are among the largest double-stranded DNA viruses, with members such as variola virus, monkeypox virus and the vaccination strain vaccinia virus (VACV). Knowledge about the structural proteins that form the viral core has remained sparse. While major core proteins have been annotated via indirect experimental evidence, their structures have remained elusive and they could not be assigned to individual core features. Hence, which proteins constitute which layers of the core, such as the palisade layer and the inner core wall, has remained enigmatic. Here we show, using a multi-modal cryo-electron microscopy (cryo-EM) approach in combination with AlphaFold molecular modeling, that trimers formed by the cleavage product of VACV protein A10 are the key component of the palisade layer. This allows us to place previously obtained descriptions of protein interactions within the core wall into perspective and to provide a detailed model of poxvirus core architecture. Importantly, we show that interactions within A10 trimers are likely generalizable over members of orthopox- and parapoxviruses.}, author = {Datler, Julia and Hansen, Jesse and Thader, Andreas and Schlögl, Alois and Bauer, Lukas W and Hodirnau, Victor-Valentin and Schur, Florian KM}, issn = {1545-9985}, journal = {Nature Structural & Molecular Biology}, keywords = {Molecular Biology, Structural Biology}, pages = {1114--1123}, publisher = {Springer Nature}, title = {{Multi-modal cryo-EM reveals trimers of protein A10 to form the palisade layer in poxvirus cores}}, doi = {10.1038/s41594-023-01201-6}, volume = {31}, year = {2024}, } @phdthesis{17850, abstract = {Understanding the relationship between a given phenotype and its underlying genotype or genotypes is one of the most pressing challenges of biology, as it lies at the heart of not only basic understanding of evolutionary theory, but also of practical applications in medicine and bioengineering. Understanding this relationship is complicated by the ubiquitous phenomenon of epistasis, wherein mutation effects are dependent on their genetic context. Fitness landscapes — representations of phenotype as a function of genotype — are being increasingly used as a tool to study the effects and interactions of thousands of mutations, but are experimentally limited to exploring a small fraction of a protein’s theoretical sequence space. Furthermore, not all regions of said sequence space are necessarily equally informative. Thus, gene selection for landscape surveys should be carefully considered in order to maximize the usable output of necessarily limited data. In this work, we analyzed the fitness landscapes of orthologous green fluorescent proteins from four different species, by systematically measuring the phenotype, fluorescence, of tens of thousands of mutant genotypes from each protein. These landscapes were highly heterogeneous, with some genes being mutationally robust and displaying epistasis only rarely, and others being highly epistatic and mutationally fragile. We used this data to train machine learning models to predict fluorescence from genotype. Although the training data contained almost exclusively genotypes with less than 3% sequence divergence from the original wild-type sequences, we were able to create novel, functional genotypes with up to 20% sequence divergence. Counterintuitively however, genes with high mutational robustness and rare epistasis were more difficult to introduce large numbers of mutations into, not less. This represents the first study of large-scale fitness landscapes of a protein family, and provides insights into how to approach future landscape surveys and their applications in novel protein design.}, author = {Gonzalez Somermeyer, Louisa}, issn = {2663-337X}, pages = {89}, publisher = {Institute of Science and Technology Austria}, title = {{Fitness landscapes of orthologous green fluorescent proteins}}, doi = {10.15479/at:ista:17850}, year = {2024}, } @article{14901, abstract = {Global services like navigation, communication, and Earth observation have increased dramatically in the 21st century due to advances in outer space industries. But as orbits become increasingly crowded with both satellites and inevitable space debris pollution, continued operations become endangered by the heightened risks of debris collisions in orbit. Kessler Syndrome is the term for when a critical threshold of orbiting debris triggers a runaway positive feedback loop of debris collisions, creating debris congestion that can render orbits unusable. As this potential tipping point becomes more widely recognized, there have been renewed calls for debris mitigation and removal. Here, we combine complex systems and social-ecological systems approaches to study how these efforts may affect space debris accumulation and the likelihood of reaching Kessler Syndrome. Specifically, we model how debris levels are affected by future launch rates, cleanup activities, and collisions between extant debris. We contextualize and interpret our dynamic model within a discussion of existing space debris governance and other social, economic, and geopolitical factors that may influence effective collective management of the orbital commons. In line with previous studies, our model finds that debris congestion may be reached in less than 200 years, though a holistic management strategy combining removal and mitigation actions can avoid such outcomes while continuing space activities. Moreover, although active debris removal may be particularly effective, the current lack of market and governance support may impede its implementation. Research into these critical dynamics and the multi-faceted variables that influence debris outcomes can support policymakers in curating impactful governance strategies and realistic transition pathways to sustaining debris-free orbits. Overall, our study is useful for communicating about space debris sustainability in policy and education settings by providing an exploration of policy portfolio options supported by a simple and clear social-ecological modeling approach.}, author = {Nomura, Keiko and Rella, Simon and Merritt, Haily and Baltussen, Mathieu and Bird, Darcy and Tjuka, Annika and Falk, Dan}, issn = {1875-0281}, journal = {International Journal of the Commons}, keywords = {Sociology and Political Science}, number = {1}, publisher = {Ubiquity Press}, title = {{Tipping points of space debris in low earth orbit}}, doi = {10.5334/ijc.1275}, volume = {18}, year = {2024}, } @phdthesis{18076, abstract = {The new era of Ge has opened up new possibilities in quantum computing. The maturity of Ge spin qubits is unquestioned, while hybrid semiconductor-superconductor Ge circuits are on track to enter the game. Gate-tunable transmons (gatemons) employing semiconductor Josephson junctions have recently emerged as building blocks for such hybrid quantum circuits. In this thesis, we present a gatemon fabricated in planar Germanium. We induce superconductivity in a two-dimensional hole gas by evaporating aluminum atop a thin spacer, which separates the superconductor from the Ge quantum well. The Josephson junction is then integrated into an Xmon circuit and capacitively coupled to a transmission line resonator. We showcase the qubit tunability in a broad frequency range with resonator and two-tone spectroscopy. Time-domain characterizations reveal energy relaxation and coherence times up to 75 ns. Our results, combined with the recent advances in the spin qubit field, pave the way towards novel hybrid and protected qubits in a group IV, CMOS-compatible material.}, author = {Sagi, Oliver}, issn = {2663-337X}, pages = {111}, publisher = {Institute of Science and Technology Austria}, title = {{Hybrid circuits on planar Germanium}}, doi = {10.15479/at:ista:18076}, year = {2024}, } @inproceedings{17170, abstract = {In this article we extend and strengthen the seminal work by Niyogi, Smale, and Weinberger on the learning of the homotopy type from a sample of an underlying space. In their work, Niyogi, Smale, and Weinberger studied samples of C² manifolds with positive reach embedded in ℝ^d. We extend their results in the following ways: - As the ambient space we consider both ℝ^d and Riemannian manifolds with lower bounded sectional curvature. - In both types of ambient spaces, we study sets of positive reach - a significantly more general setting than C² manifolds - as well as general manifolds of positive reach. - The sample P of a set (or a manifold) 𝒮 of positive reach may be noisy. We work with two one-sided Hausdorff distances - ε and δ - between P and 𝒮. We provide tight bounds in terms of ε and δ, that guarantee that there exists a parameter r such that the union of balls of radius r centred at the sample P deformation-retracts to 𝒮. We exhibit their tightness by an explicit construction. We carefully distinguish the roles of δ and ε. This is not only essential to achieve tight bounds, but also sensible in practical situations, since it allows one to adapt the bound according to sample density and the amount of noise present in the sample separately.}, author = {Attali, Dominique and Kourimska, Hana and Fillmore, Christopher D and Ghosh, Ishika and Lieutier, André and Stephenson, Elizabeth R and Wintraecken, Mathijs}, booktitle = {40th International Symposium on Computational Geometry}, isbn = {9783959773164}, issn = {1868-8969}, location = {Athens, Greece}, pages = {11:1--11:19}, publisher = {Schloss Dagstuhl - Leibniz-Zentrum für Informatik}, title = {{Tight bounds for the learning of homotopy à la Niyogi, Smale, and Weinberger for subsets of euclidean spaces and of Riemannian manifolds}}, doi = {10.4230/LIPIcs.SoCG.2024.11}, volume = {293}, year = {2024}, } @article{17202, abstract = {Gate-tunable transmons (gatemons) employing semiconductor Josephson junctions have recently emerged as building blocks for hybrid quantum circuits. In this study, we present a gatemon fabricated in planar Germanium. We induce superconductivity in a two-dimensional hole gas by evaporating aluminum atop a thin spacer, which separates the superconductor from the Ge quantum well. The Josephson junction is then integrated into an Xmon circuit and capacitively coupled to a transmission line resonator. We showcase the qubit tunability in a broad frequency range with resonator and two-tone spectroscopy. Time-domain characterizations reveal energy relaxation and coherence times up to 75 ns. Our results, combined with the recent advances in the spin qubit field, pave the way towards novel hybrid and protected qubits in a group IV, CMOS-compatible material.}, author = {Sagi, Oliver and Crippa, Alessandro and Valentini, Marco and Janik, Marian and Baghumyan, Levon and Fabris, Giorgio and Kapoor, Lucky and Hassani, Farid and Fink, Johannes M and Calcaterra, Stefano and Chrastina, Daniel and Isella, Giovanni and Katsaros, Georgios}, issn = {2041-1723}, journal = {Nature Communications}, publisher = {Springer Nature}, title = {{A gate tunable transmon qubit in planar Ge}}, doi = {10.1038/s41467-024-50763-6}, volume = {15}, year = {2024}, } @misc{17196, abstract = {This .zip File contains the data for the figures presented in the main text and supplementary material of "A gate tunable transmon qubit in planar Ge" by O.Sagi et al. The measurements were done using Qcodes. The description of the files and the instructions on opening the data can be found in the Readme. An additional Jupyter Notebook is attached that walks through the data analysis.}, author = {Sagi, Oliver}, publisher = {Institute of Science and Technology Austria}, title = {{A gate-tunable transmon in planar Ge}}, doi = {10.15479/AT:ISTA:17196}, year = {2024}, } @article{17141, abstract = {The TIR1/AFB–Aux/IAA–ARF canonical auxin signaling pathway is widely accepted to (de)active transcriptional regulation, thus controlling auxin-associated developmental processes. However, the theme of a rapid auxin response has emerged since the 2018 Auxins and Cytokinin in Plant Development conference. To date, a few signaling components have been identified to mediate both slow and rapid auxin responses, which unveils the complexity of auxin signaling.}, author = {Zhang, Zilin and Chen, Huihuang and Peng, Shuaiying and Han, Huibin}, issn = {0022-0957}, journal = {Journal of Experimental Botany}, number = {18}, publisher = {Oxford University Press}, title = {{Slow and rapid auxin responses in Arabidopsis}}, doi = {10.1093/jxb/erae246}, volume = {75}, year = {2024}, } @unpublished{18689, abstract = {Multiplexed fluorescence microscopy imaging is widely used in biomedical applications. However, simultaneous imaging of multiple fluorophores can result in spectral leaks and overlapping, which greatly degrades image quality and subsequent analysis. Existing popular spectral unmixing methods are mainly based on computational intensive linear models and the performance is heavily dependent on the reference spectra, which may greatly preclude its further applications. In this paper, we propose a deep learning-based blindly spectral unmixing method, termed AutoUnmix, to imitate the physical spectral mixing process. A tranfer learning framework is further devised to allow our AutoUnmix adapting to a variety of imaging systems without retraining the network. Our proposed method has demonstrated real-time unmixing capabilities, surpassing existing methods by up to 100-fold in terms of unmixing speed. We further validate the reconstruction performance on both synthetic datasets and biological samples. The unmixing results of AutoUnmix achieve a highest SSIM of 0.99 in both three- and four-color imaging, with nearly up to 20% higher than other popular unmixing methods. Due to the desirable property of data independency and superior blind unmixing performance, we believe AutoUnmix is a powerful tool to study the interaction process of different organelles labeled by multiple fluorophores.}, author = {Gallei, Michelle C and Truckenbrodt, Sven M and Kreuzinger, Caroline and Inumella, Syamala and Vistunou, Vitali and Sommer, Christoph M and Tavakoli, Mojtaba and Agudelo Duenas, Nathalie and Vorlaufer, Jakob and Jahr, Wiebke and Randuch, Marek and Johnson, Alexander J and Benková, Eva and Friml, Jiří and Danzl, Johann G}, booktitle = {bioRxiv}, title = {{Super-resolution expansion microscopy in plant roots}}, doi = {10.1101/2024.02.21.581330}, year = {2024}, } @phdthesis{18681, author = {Tavakoli, Mojtaba}, isbn = {978-3-99078-048-0}, issn = {2663-337X}, pages = {230}, publisher = {Institute of Science and Technology Austria}, title = {{Developing molecular and structural tools for studying brain architecture with super resolution expansion microscopy. LICONN: Molecularly-informed connectomics reconstruction with light microscopy}}, doi = {10.15479/at:ista:18681}, year = {2024}, } @unpublished{18677, abstract = {The information-processing capability of the brain’s cellular network depends on the physical wiring pattern between neurons and their molecular and functional characteristics. Mapping neurons and resolving their individual synaptic connections can be achieved by volumetric imaging at nanoscale resolution with dense cellular labeling. Light microscopy is uniquely positioned to visualize specific molecules but dense, synapse-level circuit reconstruction by light microscopy has been out of reach due to limitations in resolution, contrast, and volumetric imaging capability. Here we developed light-microscopy based connectomics (LICONN). We integrated specifically engineered hydrogel embedding and expansion with comprehensive deep-learning based segmentation and analysis of connectivity, thus directly incorporating molecular information in synapse-level brain tissue reconstructions. LICONN will allow synapse-level brain tissue phenotyping in biological experiments in a readily adoptable manner.}, author = {Tavakoli, Mojtaba and Lyudchik, Julia and Januszewski, Michał and Vistunou, Vitali and Agudelo Duenas, Nathalie and Vorlaufer, Jakob and Sommer, Christoph M and Kreuzinger, Caroline and Oliveira, Bárbara and Cenameri, Alban and Novarino, Gaia and Jain, Viren and Danzl, Johann G}, booktitle = {bioRxiv}, title = {{Light-microscopy based dense connectomic reconstruction of mammalian brain tissue}}, doi = {10.1101/2024.03.01.582884}, year = {2024}, } @phdthesis{17133, abstract = {An ideal quantum computer relies on qubits capable of performing fast gate operations and maintaining strong interconnections while preserving their quantum coherence. Since the inception of experimental eforts toward building a quantum computer, the community has faced challenges in engineering such a system. Among the various methods of implementing a quantum computer, superconducting qubits have shown fast gates close to tens of nanoseconds, with the state-of-the-art reaching a coherence of a few milliseconds. However, achieving simultaneously long lifetimes with fast qubit operations poses an inherent paradox. Qubits with high coherence require isolation from the environment, while fast operation necessitates strong coupling of the qubit. This thesis approaches this issue by proposing the idea of engineering superconducting qubits capable of transitioning between operating in a protected regime, where the qubit is completely isolated from the environment, and coupling to the communication channels as needed. In this direction, we use the geometric superinductor to scan the parameter space of rf-SQUID devices, searching for a regime where we can take the qubit protection to its extreme. This leads us to the inductively shunted transmon (IST) regime, characterized by EJ /EC ≫ 1 and EJ /EL ≫ 1, where the circuit potential exhibits a double well with a large barrier separating the local ground states of each quantum well. In this regime, although it is anticipated that the two quantum wells would be isolated from each other, we observe single fuxon tunneling between them. The interplay of the cavity photons and the fuxon transition forms a rich physical system, containing resonance conditions that allow the preparation of the fuxon ground or excited states. This enables us to study the relaxation rate of such transition and show that it can be as large as 3.6 hours. Dynamically controlling the barrier height between the two quantum wells allows for controllable coupling, which scales exponentially, for a qubit encoded in two fuxon states. The 0-π qubit is one of the very few known superconducting circuit types that ofers exponential protection from both relaxation and dephasing simultaneously. However, this qubit is not exempt from the fact that such protection comes at the expense of complex readout and control. In this thesis, we propose a way to controllably break the circuit symmetry, the key reason for the protection, to momentarily restore the ability to control and manipulate the qubit. An asymmetry in capacitances and inductances in the 0-π circuit is detrimental since they lead to coupling of the protected state to the thermally occupied parasitic mode of the circuit. However, here we try to exploit a controlled asymmetry in Josephson energies and show that this can be used as a tunable coupler between the protected states. In the future, this should allow to perform gate operations by dynamically controlling the asymmetry instead of driving the protected transition with microwave pulses. Therefore, we believe that the proposed method can make the use of protected qubits more practical in experimental realizations of quantum computing.}, author = {Hassani, Farid}, isbn = {978-3-99078-040-4}, issn = {2663-337X}, keywords = {Quantum information, Qubits, Superconducting devices}, pages = {161}, publisher = {Institute of Science and Technology Austria}, title = {{Superconducting qubits capable of dynamic switching between protected and high-speed control regimes}}, doi = {10.15479/at:ista:17133}, year = {2024}, } @article{14852, abstract = {The physical conditions giving rise to high escape fractions of ionizing radiation (LyC fesc) in star-forming galaxies – most likely protagonists of cosmic reionization – are not yet fully understood. Using the VLT/MUSE observations of ∼1400 Ly α emitters at 2.9 < z < 6.7, we compare stacked rest-frame UV spectra of candidates for LyC leakers and non-leakers selected based on their Ly α profiles. We find that the stacks of potential LyC leakers, i.e. galaxies with narrow, symmetric Ly α profiles with small peak separation, generally show (i) strong nebular O iii]λ1666, [Si iii]λ1883, and [C iii]λ1907 +C iii]λ1909 emission, indicating a high-ionization state of the interstellar medium (ISM); (ii) high equivalent widths of He iiλ1640 (∼1 − 3 Å), suggesting the presence of hard ionizing radiation fields; (iii) Si ii*λ1533 emission, revealing substantial amounts of neutral hydrogen off the line of sight; (iv) high C ivλλ1548,1550 to [C iii]λ1907 +C iii]λ1909 ratios (C iv/C iii] ≳0.75) , signalling the presence of low column density channels in the ISM. In contrast, the stacks with broad, asymmetric Ly α profiles with large peak separation show weak nebular emission lines, low He iiλ1640 equivalent widths (≲1 Å), and low C iv/C iii] (≲0.25), implying low-ionization states and high-neutral hydrogen column densities. Our results suggest that C iv/C iii] might be sensitive to the physical conditions that govern LyC photon escape, providing a promising tool for identification of ionizing sources among star-forming galaxies in the epoch of reionization.}, author = {Kramarenko, Ivan and Kerutt, J and Verhamme, A and Oesch, P A and Barrufet, L and Matthee, Jorryt J and Kusakabe, H and Goovaerts, I and Thai, T T}, issn = {1365-2966}, journal = {Monthly Notices of the Royal Astronomical Society}, keywords = {Space and Planetary Science, Astronomy and Astrophysics}, number = {4}, pages = {9853--9871}, publisher = {Oxford University Press}, title = {{Linking UV spectral properties of MUSE Ly α emitters at z ≳ 3 to Lyman continuum escape}}, doi = {10.1093/mnras/stad3853}, volume = {527}, year = {2024}, } @article{18266, abstract = {Matrix games are the most basic model in game theory, and yet robustness with respect to small perturbations of the matrix entries is not fully understood. In this paper, we introduce value positivity and uniform value positivity, two properties that refine the notion of optimality in the context of polynomially perturbed matrix games. The first concept captures how the value depends on the perturbation parameter, and the second consists of the existence of a fixed strategy that guarantees the value of the unperturbed matrix game for every sufficiently small positive parameter. We provide polynomial-time algorithms to check whether a polynomially perturbed matrix game satisfies these properties. We further provide the functional form for a parameterized optimal strategy and the value function. Finally, we translate our results to linear programming and stochastic games, where value positivity is related to the existence of robust solutions.}, author = {Chatterjee, Krishnendu and Oliu-Barton, Miquel and Saona Urmeneta, Raimundo J}, issn = {1526-5471}, journal = {Mathematics of Operations Research}, number = {4}, pages = {2433--3282}, publisher = {Institute for Operations Research and the Management Sciences}, title = {{Value-positivity for matrix games}}, doi = {10.1287/moor.2022.0332}, volume = {50}, year = {2024}, } @article{17047, abstract = {We provide a dynamical study of a model of multiplicative perturbation of a unitary matrix introduced by Fyodorov. In particular, we identify a flow of deterministic domains that bound the spectrum with high probability, separating the outlier from the typical eigenvalues at all sub-critical timescales. These results are obtained under generic assumptions on U that hold for a variety of unitary random matrix models.}, author = {Dubach, Guillaume and Reker, Jana}, issn = {2010-3271}, journal = {Random Matrices: Theory and Applications}, number = {2}, publisher = {World Scientific Publishing}, title = {{Dynamics of a rank-one multiplicative perturbation of a unitary matrix}}, doi = {10.1142/s2010326324500072}, volume = {13}, year = {2024}, } @inproceedings{18891, abstract = {Deep neural networks (DNNs) exhibit a surprising structure in their final layer known as neural collapse (NC), and a growing body of works has currently investigated the propagation of neural collapse to earlier layers of DNNs – a phenomenon called deep neural collapse (DNC). However, existing theoretical results are restricted to special cases: linear models, only two layers or binary classification. In contrast, we focus on non-linear models of arbitrary depth in multi-class classification and reveal a surprising qualitative shift. As soon as we go beyond two layers or two classes, DNC stops being optimal for the deep unconstrained features model (DUFM) – the standard theoretical framework for the analysis of collapse. The main culprit is a low-rank bias of multi-layer regularization schemes: this bias leads to optimal solutions of even lower rank than the neural collapse. We support our theoretical findings with experiments on both DUFM and real data, which show the emergence of the low-rank structure in the solution found by gradient descent.}, author = {Súkeník, Peter and Lampert, Christoph and Mondelli, Marco}, booktitle = {38th Annual Conference on Neural Information Processing Systems}, location = {Vancouver, Canada}, publisher = {Neural Information Processing Systems Foundation}, title = {{Neural collapse versus low-rank bias: Is deep neural collapse really optimal?}}, volume = {37}, year = {2024}, }