@inproceedings{9415, abstract = {Optimizing convolutional neural networks for fast inference has recently become an extremely active area of research. One of the go-to solutions in this context is weight pruning, which aims to reduce computational and memory footprint by removing large subsets of the connections in a neural network. Surprisingly, much less attention has been given to exploiting sparsity in the activation maps, which tend to be naturally sparse in many settings thanks to the structure of rectified linear (ReLU) activation functions. In this paper, we present an in-depth analysis of methods for maximizing the sparsity of the activations in a trained neural network, and show that, when coupled with an efficient sparse-input convolution algorithm, we can leverage this sparsity for significant performance gains. To induce highly sparse activation maps without accuracy loss, we introduce a new regularization technique, coupled with a new threshold-based sparsification method based on a parameterized activation function called Forced-Activation-Threshold Rectified Linear Unit (FATReLU). We examine the impact of our methods on popular image classification models, showing that most architectures can adapt to significantly sparser activation maps without any accuracy loss. Our second contribution is showing that these these compression gains can be translated into inference speedups: we provide a new algorithm to enable fast convolution operations over networks with sparse activations, and show that it can enable significant speedups for end-to-end inference on a range of popular models on the large-scale ImageNet image classification task on modern Intel CPUs, with little or no retraining cost. }, author = {Kurtz, Mark and Kopinsky, Justin and Gelashvili, Rati and Matveev, Alexander and Carr, John and Goin, Michael and Leiserson, William and Moore, Sage and Nell, Bill and Shavit, Nir and Alistarh, Dan-Adrian}, booktitle = {37th International Conference on Machine Learning, ICML 2020}, issn = {2640-3498}, location = {Online}, pages = {5533--5543}, title = {{Inducing and exploiting activation sparsity for fast neural network inference}}, volume = {119}, year = {2020}, } @article{9630, abstract = {Various kinds of data are routinely represented as discrete probability distributions. Examples include text documents summarized by histograms of word occurrences and images represented as histograms of oriented gradients. Viewing a discrete probability distribution as a point in the standard simplex of the appropriate dimension, we can understand collections of such objects in geometric and topological terms. Importantly, instead of using the standard Euclidean distance, we look into dissimilarity measures with information-theoretic justification, and we develop the theory needed for applying topological data analysis in this setting. In doing so, we emphasize constructions that enable the usage of existing computational topology software in this context.}, author = {Edelsbrunner, Herbert and Virk, Ziga and Wagner, Hubert}, issn = {1920180X}, journal = {Journal of Computational Geometry}, number = {2}, pages = {162--182}, publisher = {Carleton University}, title = {{Topological data analysis in information space}}, doi = {10.20382/jocg.v11i2a7}, volume = {11}, year = {2020}, } @inproceedings{9631, abstract = {The ability to leverage large-scale hardware parallelism has been one of the key enablers of the accelerated recent progress in machine learning. Consequently, there has been considerable effort invested into developing efficient parallel variants of classic machine learning algorithms. However, despite the wealth of knowledge on parallelization, some classic machine learning algorithms often prove hard to parallelize efficiently while maintaining convergence. In this paper, we focus on efficient parallel algorithms for the key machine learning task of inference on graphical models, in particular on the fundamental belief propagation algorithm. We address the challenge of efficiently parallelizing this classic paradigm by showing how to leverage scalable relaxed schedulers in this context. We present an extensive empirical study, showing that our approach outperforms previous parallel belief propagation implementations both in terms of scalability and in terms of wall-clock convergence time, on a range of practical applications.}, author = {Aksenov, Vitaly and Alistarh, Dan-Adrian and Korhonen, Janne}, booktitle = {Advances in Neural Information Processing Systems}, isbn = {9781713829546}, issn = {10495258}, location = {Vancouver, Canada}, pages = {22361--22372}, publisher = {Curran Associates}, title = {{Scalable belief propagation via relaxed scheduling}}, volume = {33}, year = {2020}, } @inproceedings{9632, abstract = {Second-order information, in the form of Hessian- or Inverse-Hessian-vector products, is a fundamental tool for solving optimization problems. Recently, there has been significant interest in utilizing this information in the context of deep neural networks; however, relatively little is known about the quality of existing approximations in this context. Our work examines this question, identifies issues with existing approaches, and proposes a method called WoodFisher to compute a faithful and efficient estimate of the inverse Hessian. Our main application is to neural network compression, where we build on the classic Optimal Brain Damage/Surgeon framework. We demonstrate that WoodFisher significantly outperforms popular state-of-the-art methods for oneshot pruning. Further, even when iterative, gradual pruning is allowed, our method results in a gain in test accuracy over the state-of-the-art approaches, for standard image classification datasets such as ImageNet ILSVRC. We examine how our method can be extended to take into account first-order information, as well as illustrate its ability to automatically set layer-wise pruning thresholds and perform compression in the limited-data regime. The code is available at the following link, https://github.com/IST-DASLab/WoodFisher.}, author = {Singh, Sidak Pal and Alistarh, Dan-Adrian}, booktitle = {Advances in Neural Information Processing Systems}, isbn = {9781713829546}, issn = {10495258}, location = {Vancouver, Canada}, pages = {18098--18109}, publisher = {Curran Associates}, title = {{WoodFisher: Efficient second-order approximation for neural network compression}}, volume = {33}, year = {2020}, } @inproceedings{9633, abstract = {The search for biologically faithful synaptic plasticity rules has resulted in a large body of models. They are usually inspired by – and fitted to – experimental data, but they rarely produce neural dynamics that serve complex functions. These failures suggest that current plasticity models are still under-constrained by existing data. Here, we present an alternative approach that uses meta-learning to discover plausible synaptic plasticity rules. Instead of experimental data, the rules are constrained by the functions they implement and the structure they are meant to produce. Briefly, we parameterize synaptic plasticity rules by a Volterra expansion and then use supervised learning methods (gradient descent or evolutionary strategies) to minimize a problem-dependent loss function that quantifies how effectively a candidate plasticity rule transforms an initially random network into one with the desired function. We first validate our approach by re-discovering previously described plasticity rules, starting at the single-neuron level and “Oja’s rule”, a simple Hebbian plasticity rule that captures the direction of most variability of inputs to a neuron (i.e., the first principal component). We expand the problem to the network level and ask the framework to find Oja’s rule together with an anti-Hebbian rule such that an initially random two-layer firing-rate network will recover several principal components of the input space after learning. Next, we move to networks of integrate-and-fire neurons with plastic inhibitory afferents. We train for rules that achieve a target firing rate by countering tuned excitation. Our algorithm discovers a specific subset of the manifold of rules that can solve this task. Our work is a proof of principle of an automated and unbiased approach to unveil synaptic plasticity rules that obey biological constraints and can solve complex functions.}, author = {Confavreux, Basile J and Zenke, Friedemann and Agnes, Everton J. and Lillicrap, Timothy and Vogels, Tim P}, booktitle = {Advances in Neural Information Processing Systems}, issn = {1049-5258}, location = {Vancouver, Canada}, pages = {16398--16408}, title = {{A meta-learning approach to (re)discover plasticity rules that carve a desired function into a neural network}}, volume = {33}, year = {2020}, } @unpublished{10012, abstract = {We prove that in the absence of topological changes, the notion of BV solutions to planar multiphase mean curvature flow does not allow for a mechanism for (unphysical) non-uniqueness. Our approach is based on the local structure of the energy landscape near a classical evolution by mean curvature. Mean curvature flow being the gradient flow of the surface energy functional, we develop a gradient-flow analogue of the notion of calibrations. Just like the existence of a calibration guarantees that one has reached a global minimum in the energy landscape, the existence of a "gradient flow calibration" ensures that the route of steepest descent in the energy landscape is unique and stable.}, author = {Fischer, Julian L and Hensel, Sebastian and Laux, Tim and Simon, Thilo}, booktitle = {arXiv}, title = {{The local structure of the energy landscape in multiphase mean curvature flow: weak-strong uniqueness and stability of evolutions}}, year = {2020}, } @unpublished{10022, abstract = {We consider finite-volume approximations of Fokker-Planck equations on bounded convex domains in R^d and study the corresponding gradient flow structures. We reprove the convergence of the discrete to continuous Fokker-Planck equation via the method of Evolutionary Γ-convergence, i.e., we pass to the limit at the level of the gradient flow structures, generalising the one-dimensional result obtained by Disser and Liero. The proof is of variational nature and relies on a Mosco convergence result for functionals in the discrete-to-continuum limit that is of independent interest. Our results apply to arbitrary regular meshes, even though the associated discrete transport distances may fail to converge to the Wasserstein distance in this generality.}, author = {Forkert, Dominik L and Maas, Jan and Portinale, Lorenzo}, booktitle = {arXiv}, pages = {33}, title = {{Evolutionary Γ-convergence of entropic gradient flow structures for Fokker-Planck equations in multiple dimensions}}, year = {2020}, } @inproceedings{10328, abstract = {We discus noise channels in coherent electro-optic up-conversion between microwave and optical fields, in particular due to optical heating. We also report on a novel configuration, which promises to be flexible and highly efficient.}, author = {Lambert, Nicholas J. and Mobassem, Sonia and Rueda Sanchez, Alfredo R and Schwefel, Harald G.L.}, booktitle = {OSA Quantum 2.0 Conference}, isbn = {9-781-5575-2820-9}, location = {Washington, DC, United States}, publisher = {Optica Publishing Group}, title = {{New designs and noise channels in electro-optic microwave to optical up-conversion}}, doi = {10.1364/QUANTUM.2020.QTu8A.1}, year = {2020}, } @inproceedings{10556, abstract = {In this paper, we present the first Asynchronous Distributed Key Generation (ADKG) algorithm which is also the first distributed key generation algorithm that can generate cryptographic keys with a dual (f,2f+1)-threshold (where f is the number of faulty parties). As a result, using our ADKG we remove the trusted setup assumption that the most scalable consensus algorithms make. In order to create a DKG with a dual (f,2f+1)- threshold we first answer in the affirmative the open question posed by Cachin et al. [7] on how to create an Asynchronous Verifiable Secret Sharing (AVSS) protocol with a reconstruction threshold of f+1 95%). (Table S2). All pQTLs associated with inflammatory biomarker levels from ordinary least squares regression model (P < 7.14 × 10− 10). (Table S3). Summary of lambda values relating to ordinary least squares GWAS and EWAS performed on inflammatory protein levels (n = 70) in Lothian Birth Cohort 1936 study. (Table S4). Conditionally significant pQTLs associated with inflammatory biomarker levels from ordinary least squares regression model (P < 7.14 × 10− 10). (Table S5). Comparison of variance explained by ordinary least squares and Bayesian penalised regression models for concordantly identified SNPs. (Table S6). Estimate of heritability for blood protein levels as well as proportion of variance explained attributable to different prior mixtures. (Table S7). Comparison of heritability estimates from Ahsan et al. (maximum likelihood) and Hillary et al. (Bayesian penalised regression). (Table S8). List of concordant SNPs identified by linear model and Bayesian penalised regression and whether they have been previously identified as eQTLs. (Table S9). Bayesian tests of colocalisation for cis pQTLs and cis eQTLs. (Table S10). Sherlock algorithm: Genes whose expression are putatively associated with circulating inflammatory proteins that harbour pQTLs. (Table S11). CpGs associated with inflammatory protein biomarkers as identified by Bayesian model (Bayesian model; Posterior Inclusion Probability > 95%). (Table S12). CpGs associated with inflammatory protein biomarkers as identified by linear model (limma) at P < 5.14 × 10− 10. (Table S13). CpGs associated with inflammatory protein biomarkers as identified by mixed linear model (OSCA) at P < 5.14 × 10− 10. (Table S14). Estimate of variance explained for blood protein levels by DNA methylation as well as proportion of explained attributable to different prior mixtures - BayesR+. (Table S15). Comparison of variance in protein levels explained by genome-wide DNA methylation data by mixed linear model (OSCA) and Bayesian penalised regression model (BayesR+). (Table S16). Variance in circulating inflammatory protein biomarker levels explained by common genetic and methylation data (joint and conditional estimates from BayesR+). Ordered by combined variance explained by genetic and epigenetic data - smallest to largest. Significant results from t-tests comparing distributions for variance explained by methylation or genetics alone versus combined estimate are emboldened. (Table S17). Genetic and epigenetic factors identified by BayesR+ when conditioning on all SNPs and CpGs together. (Table S18). Mendelian Randomisation analyses to assess whether proteins with concordantly identified genetic signals are causally associated with Alzheimer’s disease risk. (Table S19).}, author = {Hillary, Robert F. and Trejo-Banos, Daniel and Kousathanas, Athanasios and McCartney, Daniel L. and Harris, Sarah E. and Stevenson, Anna J. and Patxot, Marion and Ojavee, Sven Erik and Zhang, Qian and Liewald, David C. and Ritchie, Craig W. and Evans, Kathryn L. and Tucker-Drob, Elliot M. and Wray, Naomi R. and McRae, Allan F. and Visscher, Peter M. and Deary, Ian J. and Robinson, Matthew Richard and Marioni, Riccardo E. }, publisher = {Springer Nature}, title = {{Additional file 2 of multi-method genome- and epigenome-wide studies of inflammatory protein levels in healthy older adults}}, doi = {10.6084/m9.figshare.12629697.v1}, year = {2020}, } @misc{9708, abstract = {This research data supports 'Hard antinodal gap revealed by quantum oscillations in the pseudogap regime of underdoped high-Tc superconductors'. A Readme file for plotting each figure is provided.}, author = {Hartstein, Mate and Hsu, Yu-Te and Modic, Kimberly A and Porras, Juan and Loew, Toshinao and Le Tacon, Matthieu and Zuo, Huakun and Wang, Jinhua and Zhu, Zengwei and Chan, Mun and McDonald, Ross and Lonzarich, Gilbert and Keimer, Bernhard and Sebastian, Suchitra and Harrison, Neil}, publisher = {Apollo - University of Cambridge}, title = {{Accompanying dataset for 'Hard antinodal gap revealed by quantum oscillations in the pseudogap regime of underdoped high-Tc superconductors'}}, doi = {10.17863/cam.50169}, year = {2020}, } @misc{9713, abstract = {Additional analyses of the trajectories}, author = {Gupta, Chitrak and Khaniya, Umesh and Chan, Chun Kit and Dehez, Francois and Shekhar, Mrinal and Gunner, M.R. and Sazanov, Leonid A and Chipot, Christophe and Singharoy, Abhishek}, publisher = {American Chemical Society }, title = {{Supporting information}}, doi = {10.1021/jacs.9b13450.s001}, year = {2020}, } @unpublished{9750, abstract = {Tension of the actomyosin cell cortex plays a key role in determining cell-cell contact growth and size. The level of cortical tension outside of the cell-cell contact, when pulling at the contact edge, scales with the total size to which a cell-cell contact can grow1,2. Here we show in zebrafish primary germ layer progenitor cells that this monotonic relationship only applies to a narrow range of cortical tension increase, and that above a critical threshold, contact size inversely scales with cortical tension. This switch from cortical tension increasing to decreasing progenitor cell-cell contact size is caused by cortical tension promoting E-cadherin anchoring to the actomyosin cytoskeleton, thereby increasing clustering and stability of E-cadherin at the contact. Once tension-mediated E-cadherin stabilization at the contact exceeds a critical threshold level, the rate by which the contact expands in response to pulling forces from the cortex sharply drops, leading to smaller contacts at physiologically relevant timescales of contact formation. Thus, the activity of cortical tension in expanding cell-cell contact size is limited by tension stabilizing E-cadherin-actin complexes at the contact.}, author = {Slovakova, Jana and Sikora, Mateusz K and Caballero Mancebo, Silvia and Krens, Gabriel and Kaufmann, Walter and Huljev, Karla and Heisenberg, Carl-Philipp J}, booktitle = {bioRxiv}, pages = {41}, publisher = {Cold Spring Harbor Laboratory}, title = {{Tension-dependent stabilization of E-cadherin limits cell-cell contact expansion}}, doi = {10.1101/2020.11.20.391284}, year = {2020}, } @misc{9776, author = {Grah, Rok and Friedlander, Tamar}, publisher = {Public Library of Science}, title = {{Supporting information}}, doi = {10.1371/journal.pcbi.1007642.s001}, year = {2020}, } @misc{9777, author = {Grah, Rok and Friedlander, Tamar}, publisher = {Public Library of Science}, title = {{Maximizing crosstalk}}, doi = {10.1371/journal.pcbi.1007642.s002}, year = {2020}, } @misc{9779, author = {Grah, Rok and Friedlander, Tamar}, publisher = {Public Library of Science}, title = {{Distribution of crosstalk values}}, doi = {10.1371/journal.pcbi.1007642.s003}, year = {2020}, } @misc{9780, abstract = {PADREV : 4,4'-dimethoxy[1,1'-biphenyl]-2,2',5,5'-tetrol Space Group: C 2 (5), Cell: a 24.488(16)Å b 5.981(4)Å c 3.911(3)Å, α 90° β 91.47(3)° γ 90°}, author = {Schlemmer, Werner and Nothdurft, Philipp and Petzold, Alina and Riess, Gisbert and Frühwirt, Philipp and Schmallegger, Max and Gescheidt-Demner, Georg and Fischer, Roland and Freunberger, Stefan Alexander and Kern, Wolfgang and Spirk, Stefan}, publisher = {CCDC}, title = {{CCDC 1991959: Experimental Crystal Structure Determination}}, doi = {10.5517/ccdc.csd.cc24vsrk}, year = {2020}, } @article{9781, abstract = {We consider the Pekar functional on a ball in ℝ3. We prove uniqueness of minimizers, and a quadratic lower bound in terms of the distance to the minimizer. The latter follows from nondegeneracy of the Hessian at the minimum.}, author = {Feliciangeli, Dario and Seiringer, Robert}, issn = {1095-7154}, journal = {SIAM Journal on Mathematical Analysis}, keywords = {Applied Mathematics, Computational Mathematics, Analysis}, number = {1}, pages = {605--622}, publisher = {Society for Industrial & Applied Mathematics }, title = {{Uniqueness and nondegeneracy of minimizers of the Pekar functional on a ball}}, doi = {10.1137/19m126284x}, volume = {52}, year = {2020}, } @misc{9798, abstract = {Fitness interactions between mutations can influence a population’s evolution in many different ways. While epistatic effects are difficult to measure precisely, important information is captured by the mean and variance of log fitnesses for individuals carrying different numbers of mutations. We derive predictions for these quantities from a class of simple fitness landscapes, based on models of optimizing selection on quantitative traits. We also explore extensions to the models, including modular pleiotropy, variable effect sizes, mutational bias and maladaptation of the wild type. We illustrate our approach by reanalysing a large dataset of mutant effects in a yeast snoRNA. Though characterized by some large epistatic effects, these data give a good overall fit to the non-epistatic null model, suggesting that epistasis might have limited influence on the evolutionary dynamics in this system. We also show how the amount of epistasis depends on both the underlying fitness landscape and the distribution of mutations, and so is expected to vary in consistent ways between new mutations, standing variation and fixed mutations.}, author = {Fraisse, Christelle and Welch, John J.}, publisher = {Royal Society of London}, title = {{Simulation code for Fig S2 from the distribution of epistasis on simple fitness landscapes}}, doi = {10.6084/m9.figshare.7957472.v1}, year = {2020}, } @misc{9799, abstract = {Fitness interactions between mutations can influence a population’s evolution in many different ways. While epistatic effects are difficult to measure precisely, important information is captured by the mean and variance of log fitnesses for individuals carrying different numbers of mutations. We derive predictions for these quantities from a class of simple fitness landscapes, based on models of optimizing selection on quantitative traits. We also explore extensions to the models, including modular pleiotropy, variable effect sizes, mutational bias and maladaptation of the wild type. We illustrate our approach by reanalysing a large dataset of mutant effects in a yeast snoRNA. Though characterized by some large epistatic effects, these data give a good overall fit to the non-epistatic null model, suggesting that epistasis might have limited influence on the evolutionary dynamics in this system. We also show how the amount of epistasis depends on both the underlying fitness landscape and the distribution of mutations, and so is expected to vary in consistent ways between new mutations, standing variation and fixed mutations.}, author = {Fraisse, Christelle and Welch, John J.}, publisher = {Royal Society of London}, title = {{Simulation code for Fig S1 from the distribution of epistasis on simple fitness landscapes}}, doi = {10.6084/m9.figshare.7957469.v1}, year = {2020}, }