@article{14793, abstract = {Superconductor/semiconductor hybrid devices have attracted increasing interest in the past years. Superconducting electronics aims to complement semiconductor technology, while hybrid architectures are at the forefront of new ideas such as topological superconductivity and protected qubits. In this work, we engineer the induced superconductivity in two-dimensional germanium hole gas by varying the distance between the quantum well and the aluminum. We demonstrate a hard superconducting gap and realize an electrically and flux tunable superconducting diode using a superconducting quantum interference device (SQUID). This allows to tune the current phase relation (CPR), to a regime where single Cooper pair tunneling is suppressed, creating a sin(2y) CPR. Shapiro experiments complement this interpretation and the microwave drive allows to create a diode with ≈ 100% efficiency. The reported results open up the path towards integration of spin qubit devices, microwave resonators and (protected) superconducting qubits on the same silicon technology compatible platform.}, author = {Valentini, Marco and Sagi, Oliver and Baghumyan, Levon and de Gijsel, Thijs and Jung, Jason and Calcaterra, Stefano and Ballabio, Andrea and Aguilera Servin, Juan L and Aggarwal, Kushagra and Janik, Marian and Adletzberger, Thomas and Seoane Souto, Rubén and Leijnse, Martin and Danon, Jeroen and Schrade, Constantin and Bakkers, Erik and Chrastina, Daniel and Isella, Giovanni and Katsaros, Georgios}, issn = {2041-1723}, journal = {Nature Communications}, publisher = {Springer Nature}, title = {{Parity-conserving Cooper-pair transport and ideal superconducting diode in planar germanium}}, doi = {10.1038/s41467-023-44114-0}, volume = {15}, year = {2024}, } @article{18902, author = {Zagorski, Marcin and Brandenberg, Nathalie and Lutolf, Matthias and Tkačik, Gašper and Bollenbach, Mark Tobias and Briscoe, James and Kicheva, Anna}, issn = {2041-1723}, journal = {Nature Communications}, publisher = {Springer Nature}, title = {{Assessing the precision of morphogen gradients in neural tube development}}, doi = {10.1038/s41467-024-45148-8}, volume = {15}, year = {2024}, } @article{15320, abstract = {Josephson diodes are superconducting elements that show an asymmetry in the critical current depending on the direction of the current. Here, we theoretically explore how an alternating current bias can tune the response of such a diode. We show that for slow driving there is always a regime where the system can only carry zero-voltage dc current in one direction, thus effectively behaving as an ideal Josephson diode. Under fast driving, the diode efficiency is also tunable, although the ideal regime cannot be reached in this case. We also investigate the residual dissipation due to the time-dependent current bias and show that it remains small. All our conclusions are solely based on the critical current asymmetry of the junction, and are thus compatible with any Josephson diode.}, author = {Seoane Souto, Rubén and Leijnse, Martin and Schrade, Constantin and Valentini, Marco and Katsaros, Georgios and Danon, Jeroen}, issn = {2643-1564}, journal = {Physical Review Research}, number = {2}, publisher = {American Physical Society}, title = {{Tuning the Josephson diode response with an ac current}}, doi = {10.1103/PhysRevResearch.6.L022002}, volume = {6}, year = {2024}, } @article{17389, abstract = {The potential of Si and SiGe-based devices for the scaling of quantum circuits is tainted by device variability. Each device needs to be tuned to operation conditions and each device realisation requires a different tuning protocol. We demonstrate that it is possible to automate the tuning of a 4-gate Si FinFET, a 5-gate GeSi nanowire and a 7-gate Ge/SiGe heterostructure double quantum dot device from scratch with the same algorithm. We achieve tuning times of 30, 10, and 92 min, respectively. The algorithm also provides insight into the parameter space landscape for each of these devices, allowing for the characterization of the regions where double quantum dot regimes are found. These results show that overarching solutions for the tuning of quantum devices are enabled by machine learning.}, author = {Severin, B. and Lennon, D. T. and Camenzind, L. C. and Vigneau, F. and Fedele, F. and Jirovec, Daniel and Ballabio, A. and Chrastina, D. and Isella, G. and de Kruijf, M. and Carballido, M. J. and Svab, S. and Kuhlmann, A. V. and Geyer, S. and Froning, F. N. M. and Moon, H. and Osborne, M. A. and Sejdinovic, D. and Katsaros, Georgios and Zumbühl, D. M. and Briggs, G. A. D. and Ares, N.}, issn = {2045-2322}, journal = {Scientific Reports}, publisher = {Springer Nature}, title = {{Cross-architecture tuning of silicon and SiGe-based quantum devices using machine learning}}, doi = {10.1038/s41598-024-67787-z}, volume = {14}, year = {2024}, } @article{18602, abstract = {Semiconductor quantum dots (QDs) in planar germanium (Ge) heterostructures have emerged as front-runners for future hole-based quantum processors. Here, we present strong coupling between a hole charge qubit, defined in a double quantum dot (DQD) in planar Ge, and microwave photons in a high-impedance (Zr = 1.3 kΩ) resonator based on an array of superconducting quantum interference devices (SQUIDs). Our investigation reveals vacuum-Rabi splittings with coupling strengths up to g0/2π = 260 MHz, and a cooperativity of C ~ 100, dependent on DQD tuning. Furthermore, utilizing the frequency tunability of our resonator, we explore the quenched energy splitting associated with strong Coulomb correlation effects in Ge QDs. The observed enhanced coherence of the strongly correlated excited state signals the presence of distinct symmetries within related spin functions, serving as a precursor to the strong coupling between photons and spin-charge hybrid qubits in planar Ge. This work paves the way towards coherent quantum connections between remote hole qubits in planar Ge, required to scale up hole-based quantum processors.}, author = {De Palma, Franco and Oppliger, Fabian and Jang, Wonjin and Bosco, Stefano and Janik, Marian and Calcaterra, Stefano and Katsaros, Georgios and Isella, Giovanni and Loss, Daniel and Scarlino, Pasquale}, issn = {2041-1723}, journal = {Nature Communications}, publisher = {Springer Nature}, title = {{Strong hole-photon coupling in planar Ge for probing charge degree and strongly correlated states}}, doi = {10.1038/s41467-024-54520-7}, volume = {15}, year = {2024}, } @article{18653, abstract = {Charge sensing is a sensitive technique for probing quantum devices, of particular importance for spin-qubit readout. To achieve good readout sensitivities, the proximity of the charge sensor to the device to be measured is a necessity. However, this proximity also means that the operation of the device affects, in turn, the sensor tuning and ultimately the readout sensitivity. We present an approach for compensating for this crosstalk effect allowing for the gate voltages of the measured device to be swept in a 1-V × 1-V window while maintaining a sensor configuration chosen by a Bayesian optimizer. Our algorithm will hopefully be a major contribution to the suite of fully automated solutions required for the operation of large quantum device architectures.}, author = {Hickie, Joseph and Van Straaten, Barnaby and Fedele, Federico and Jirovec, Daniel and Ballabio, Andrea and Chrastina, Daniel and Isella, Giovanni and Katsaros, Georgios and Ares, Natalia}, issn = {2331-7019}, journal = {Physical Review Applied}, number = {6}, publisher = {American Physical Society}, title = {{Automated long-range compensation of an rf quantum dot sensor}}, doi = {10.1103/PhysRevApplied.22.064026}, volume = {22}, year = {2024}, } @misc{23037, abstract = {The data for all figures in the main text can be found in csv files in ASCII format in the corresponing folders. For Figures 4-6, the panels are numbered from top to bottom.}, author = {De Palma, Franco and Oppliger, Fabian and Jang, Wonjin and Bosco, Stefano and Janík, Marián and Calcaterra, Stefano and Katsaros, Georgios and Isella, Giovanni and Loss, Daniel and Scarlino, Pasquale}, publisher = {Zenodo}, title = {{Online data of "Strong hole-photon coupling in planar Ge for probing charge degree and strongly correlated states"}}, doi = {10.5281/zenodo.13935166}, year = {2024}, } @article{15018, abstract = {The epitaxial growth of a strained Ge layer, which is a promising candidate for the channel material of a hole spin qubit, has been demonstrated on 300 mm Si wafers using commercially available Si0.3Ge0.7 strain relaxed buffer (SRB) layers. The assessment of the layer and the interface qualities for a buried strained Ge layer embedded in Si0.3Ge0.7 layers is reported. The XRD reciprocal space mapping confirmed that the reduction of the growth temperature enables the 2-dimensional growth of the Ge layer fully strained with respect to the Si0.3Ge0.7. Nevertheless, dislocations at the top and/or bottom interface of the Ge layer were observed by means of electron channeling contrast imaging, suggesting the importance of the careful dislocation assessment. The interface abruptness does not depend on the selection of the precursor gases, but it is strongly influenced by the growth temperature which affects the coverage of the surface H-passivation. The mobility of 2.7 × 105 cm2/Vs is promising, while the low percolation density of 3 × 1010 /cm2 measured with a Hall-bar device at 7 K illustrates the high quality of the heterostructure thanks to the high Si0.3Ge0.7 SRB quality.}, author = {Shimura, Yosuke and Godfrin, Clement and Hikavyy, Andriy and Li, Roy and Aguilera Servin, Juan L and Katsaros, Georgios and Favia, Paola and Han, Han and Wan, Danny and de Greve, Kristiaan and Loo, Roger}, issn = {1369-8001}, journal = {Materials Science in Semiconductor Processing}, keywords = {Mechanical Engineering, Mechanics of Materials, Condensed Matter Physics, General Materials Science}, number = {5}, publisher = {Elsevier}, title = {{Compressively strained epitaxial Ge layers for quantum computing applications}}, doi = {10.1016/j.mssp.2024.108231}, volume = {174}, year = {2024}, } @misc{14705, abstract = {Since the commercialization of brine shrimp (genus Artemia) in the 1950s, this lineage, and in particular the model species Artemia franciscana, has been the subject of extensive research. However, our understanding of the genetic mechanisms underlying various aspects of their reproductive biology, including sex determination, are still lacking. This is partly due to the scarcity of genomic resources for Artemia species and crustaceans in general. Here, we present a chromosome-level genome assembly of Artemia franciscana (Kellogg 1906), from the Great Salt Lake, USA. The genome is 1GB, and the majority of the genome (81%) is scaffolded into 21 linkage groups using a previously published high-density linkage map. We performed coverage and FST analyses using male and female genomic and transcriptomic reads to quantify the extent of differentiation between the Z and W chromosomes. Additionally, we quantified the expression levels in male and female heads and gonads and found further evidence for dosage compensation in this species.}, author = {Elkrewi, Marwan N}, keywords = {sex chromosome evolution, genome assembly, dosage compensation}, publisher = {Institute of Science and Technology Austria}, title = {{Data from "Chromosome-level assembly of Artemia franciscana sheds light on sex-chromosome differentiation"}}, doi = {10.15479/AT:ISTA:14705}, year = {2024}, } @inproceedings{17214, abstract = {Current numerical algorithms for simulating friction fall in one of two camps: smooth solvers sacrifice the stable treatment of static friction in exchange for fast convergence, and non-smooth solvers accurately compute friction at convergence rates that are often prohibitive for large graphics applications. We introduce a novel bridge between these two ideas that computes static and dynamic friction stably and efficiently. Our key idea is to convert the highly constrained non-smooth problem into an unconstrained smooth problem using logarithmic barriers that converges to the exact solution as accuracy increases. We phrase the problem as an interior point primal-dual problem that can be solved efficiently with Newton iteration. We observe quadratic convergence despite the non-smooth nature of the original problem, and our method is well-suited for large systems of tightly packed objects with many contact points. We demonstrate the efficacy of our method with stable piles of grains and stacks of objects, complex granular flows, and robust interlocking assemblies of rigid bodies.}, author = {Chen, Yi-Lu and Ly, Mickaël and Wojtan, Christopher J}, booktitle = {Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers '24}, isbn = {9798400705250}, keywords = {physical simulation, frictional contact, rigid body mechanics, non-smooth dynamics}, location = {Denver, United States}, publisher = {Association for Computing Machinery}, title = {{Primal-dual non-smooth friction for rigid body animation}}, doi = {10.1145/3641519.3657485}, year = {2024}, } @inproceedings{17426, abstract = {The robustness of neural networks against input perturbations with bounded magnitude represents a serious concern in the deployment of deep learning models in safety-critical systems. Recently, the scientific community has focused on enhancing certifiable robustness guarantees by crafting 1-Lipschitz neural networks that leverage Lipschitz bounded dense and convolutional layers. Although different methods have been proposed in the literature to achieve this goal, understanding the performance of such methods is not straightforward, since different metrics can be relevant (e.g., training time, memory usage, accuracy, certifiable robustness) for different applications. For this reason, this work provides a thorough theoretical and empirical comparison between methods by evaluating them in terms of memory usage, speed, and certifiable robust accuracy. The paper also provides some guidelines and recommendations to support the user in selecting the methods that work best depending on the available resources. We provide code at https://github.com/berndprach/1LipschitzLayersCompared.}, author = {Prach, Bernd and Brau, Fabio and Buttazzo, Giorgio and Lampert, Christoph}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, location = {Seattle, WA, United States}, pages = {24574--24583}, publisher = {Computer Vision Foundation}, title = {{1-Lipschitz layers compared: Memory, speed, and certifiable robustness}}, doi = {10.1109/CVPR52733.2024.02320}, year = {2024}, } @unpublished{18874, abstract = {Despite extensive research since the community learned about adversarial examples 10 years ago, we still do not know how to train high-accuracy classifiers that are guaranteed to be robust to small perturbations of their inputs. Previous works often argued that this might be because no classifier exists that is robust and accurate at the same time. However, in computer vision this assumption does not match reality where humans are usually accurate and robust on most tasks of interest. We offer an alternative explanation and show that in certain settings robust generalization is only possible with unrealistically large amounts of data. More precisely we find a setting where a robust classifier exists, it is easy to learn an accurate classifier, yet it requires an exponential amount of data to learn a robust classifier. Based on this theoretical result, we explore how well robust classifiers generalize on datasets such as CIFAR-10. We come to the conclusion that on this datasets, the limitation of current robust models also lies in the generalization, and that they require a lot of data to do well on the test set. We also show that the problem is not in the expressiveness or generalization capabilities of current architectures, and that there are low magnitude features in the data which are useful for non-robust generalization but are not available for robust classifiers.}, author = {Prach, Bernd and Lampert, Christoph}, booktitle = {arXiv}, title = {{Intriguing properties of robust classification}}, doi = {10.48550/arXiv.2412.04245}, year = {2024}, } @unpublished{17361, abstract = {We present symplectic structures on the shape space of unparameterized space curves that generalize the classical Marsden-Weinstein structure. Our method integrates the Liouville 1-form of the Marsden-Weinstein structure with Riemannian structures that have been introduced in mathematical shape analysis. We also derive Hamiltonian vector fields for several classical Hamiltonian functions with respect to these new symplectic structures.}, author = {Bauer, Martin and Ishida, Sadashige and Michor, Peter W.}, booktitle = {arXiv}, keywords = {space of space curves, symplectic stuctures}, title = {{Symplectic structures on the space of space curves}}, doi = {10.48550/arXiv.2407.19908}, year = {2024}, } @article{17219, abstract = {We introduce a multi-material non-manifold mesh-based surface tracking algorithm that converts self-intersections into topological changes. Our algorithm generalizes prior work on manifold surface tracking with topological changes: it preserves surface features like mesh-based methods, and it robustly handles topological changes like level set methods. Our method also offers improved efficiency and robustness over the state of the art. We demonstrate the effectiveness of the approach on a range of examples, including complex soap film simulations with thousands of interacting bubbles, and boolean unions of non-manifold meshes consisting of millions of triangles.}, author = {Synak, Peter and Kalinov, Aleksei and Strugaru, Irina-Malina and Etemadihaghighi, Arian and Yang, Huidong and Wojtan, Christopher J}, issn = {1557-7368}, journal = {ACM Transactions on Graphics}, keywords = {surface tracking, topology change, non- manifold meshes, multi-material flows, solid modeling}, number = {4}, publisher = {Association for Computing Machinery}, title = {{Multi-material mesh-based surface tracking with implicit topology changes}}, doi = {10.1145/3658223}, volume = {43}, year = {2024}, } @phdthesis{18301, abstract = {Physics simulation in computer graphics can bring triangle meshes into topologically invalid states. The method in this thesis contributed to Heiss-Synak* and Kalinov* et al. [2024] who devised a non-manifold hybrid surface tracker—a surface tracker that repairs explicit non-manifold triangle meshes with the help of the implicit domain. Specifically, this thesis provides an algorithm for filling the holes that are left after removing problematic parts of the mesh.}, author = {Etemadihaghighi, Arian}, issn = {2791-4585}, keywords = {surface tracking, non-manifold, hole-filling, topology change, multi-material, solid-modeling}, pages = {39}, publisher = {Institute of Science and Technology Austria}, title = {{Filling the holes of non-manifold self-intersecting meshes for implicit topology changes in surface tracking}}, doi = {10.15479/at:ista:18301}, year = {2024}, } @article{12846, abstract = {We present a formula for the signed area of a spherical polygon via prequantization. In contrast to the traditional formula based on the Gauss-Bonnet theorem that requires measuring angles, the new formula mimics Green's theorem and is applicable to a wider range of degenerate spherical curves and polygons.}, author = {Chern, Albert and Ishida, Sadashige}, issn = {2470-6566}, journal = {SIAM Journal on Applied Algebra and Geometry}, number = {3}, pages = {782--796}, publisher = {Society for Industrial and Applied Mathematics}, title = {{Area formula for spherical polygons via prequantization}}, doi = {10.1137/23M1565255}, volume = {8}, year = {2024}, } @phdthesis{17485, abstract = {Large language models (LLMs) have made tremendous progress in the past few years, from being able to generate coherent text to matching or surpassing humans in a wide variety of creative, knowledge or reasoning tasks. Much of this can be attributed to massively increased scale, both in the size of the model as well as the amount of training data, from 100s of millions to 100s of billions, or even trillions. This trend is expected to continue, which, although exciting, also raises major practical concerns. Already today's 100+ billion parameter LLMs require top-of-the-line hardware just to run. Hence, it is clear that sustaining these developments will require significant efficiency advances. Historically, one of the most practical ways of improving model efficiency has been compression, especially in the form of sparsity or quantization. While this has been studied extensively in the past, existing accurate methods are all designed for models around 100 million parameters; scaling them up to ones literally 1000x larger is highly challenging. In this thesis, we introduce a new unified sparsification and quantization approach OBC, which through additional algorithmic enhancements leads to GPTQ and SparseGPT, the first techniques fast and accurate enough to compress 100+ billion parameter models to 4- or even 3-bit precision and 50% weight-sparsity, respectively. Additionally, we show how weight-only quantizion does not just bring space savings but also up to 4.5x faster generation speed, via custom GPU kernels. In fact, we show for the first time that it is possible to develop an FP16 times INT4 mixed-precision matrix multiplication kernel, called Marlin, which comes close to simultaneously maximizing both memory and compute utilization, making weight-only quantization highly practical even for multi-user serving. Further, we demonstrate that GPTQ can be scaled to widely overparametrized trillion-parameter models, where extreme sub-1-bit compression rates can be achieved without any inference slow-down, by co-designing a bespoke entropy coding scheme together with an efficient kernel. Finally, we also study compression from the perspective of someone with access to massive amounts of compute resources for training large models completely from scratch. Here the key questions evolve around the joint scaling behavior between compression, model size, and amount of training data used. Based on extensive experimental results for both vision and text models, we introduce the first scaling law which accurately captures the relationship between weight-sparsity, number of non-zero weights and data. This further allows us to characterize the optimal sparsity, which we find to increase the longer a fixed cost model is being trained. Overall, this thesis presents contributions to three different angles of large model efficiency: affordable but accurate algorithms, highly efficient systems implementations, and fundamental scaling laws for compressed training.}, author = {Frantar, Elias}, issn = {2663-337X}, pages = {129}, publisher = {Institute of Science and Technology Austria}, title = {{Compressing large neural networks: Algorithms, systems and scaling laws}}, doi = {10.15479/at:ista:17485}, year = {2024}, } @inproceedings{18062, abstract = {We explore the impact of parameter sparsity on the scaling behavior of Transformers trained on massive datasets (i.e., "foundation models"), in both vision and language domains. In this setting, we identify the first scaling law describing the relationship between weight sparsity, number of non-zero parameters, and amount of training data, which we validate empirically across model and data scales; on ViT/JFT-4B and T5/C4. These results allow us to characterize the "optimal sparsity", the sparsity level which yields the best performance for a given effective model size and training budget. For a fixed number of non-zero parameters, we identify that the optimal sparsity increases with the amount of data used for training. We also extend our study to different sparsity structures (such as the hardware-friendly n:m pattern) and strategies (such as starting from a pretrained dense model). Our findings shed light on the power and limitations of weight sparsity across various parameter and computational settings, offering both theoretical understanding and practical implications for leveraging sparsity towards computational efficiency improvements. We provide pruning and scaling law fitting code at: github.com/google-research/jaxpruner/tree/main/jaxpruner/projects/bigsparse.}, author = {Frantar, Elias and Ruiz, Carlos Riquelme and Houlsby, Neil and Alistarh, Dan-Adrian and Evci, Utku}, booktitle = {The Twelfth International Conference on Learning Representations}, location = {Vienna, Austria}, title = {{Scaling laws for sparsely-connected foundation models}}, year = {2024}, } @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}, }