@inproceedings{10673, abstract = {We propose a neural information processing system obtained by re-purposing the function of a biological neural circuit model to govern simulated and real-world control tasks. Inspired by the structure of the nervous system of the soil-worm, C. elegans, we introduce ordinary neural circuits (ONCs), defined as the model of biological neural circuits reparameterized for the control of alternative tasks. We first demonstrate that ONCs realize networks with higher maximum flow compared to arbitrary wired networks. We then learn instances of ONCs to control a series of robotic tasks, including the autonomous parking of a real-world rover robot. For reconfiguration of the purpose of the neural circuit, we adopt a search-based optimization algorithm. Ordinary neural circuits perform on par and, in some cases, significantly surpass the performance of contemporary deep learning models. ONC networks are compact, 77% sparser than their counterpart neural controllers, and their neural dynamics are fully interpretable at the cell-level.}, author = {Hasani, Ramin and Lechner, Mathias and Amini, Alexander and Rus, Daniela and Grosu, Radu}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, issn = {2640-3498}, location = {Virtual}, pages = {4082--4093}, title = {{A natural lottery ticket winner: Reinforcement learning with ordinary neural circuits}}, year = {2020}, } @inproceedings{10693, abstract = {High quality graphene heterostructures host an array of fractional quantum Hall isospin ferromagnets with diverse spin and valley orders. While a variety of phase transitions have been observed, disentangling the isospin phase diagram of these states is hampered by the absence of direct probes of spin and valley order. I will describe nonlocal transport measurements based on launching spin waves from a gate defined lateral heterojunction, performed in ultra-clean Corbino geometry graphene devices. At high magnetic fields, we find that the spin-wave transport signal is detected in all FQH states between ν = 0 and 1; however, between ν = 1 and 2 only odd numerator FQH states show finite nonlocal transport, despite the identical ground state spin polarizations in odd- and even numerator states. The results reveal that the neutral spin-waves are both spin and sublattice polarized making them a sensitive probe of ground state sublattice structure. Armed with this understanding, we use nonlocal transport signal to a magnetic field tuned isospin phase transition, showing that the emergent even denominator state at ν = 1/2 in monolayer graphene is indeed a multicomponent state featuring equal populations on each sublattice.}, author = {Zhou, Haoxin and Polshyn, Hryhoriy and Tanaguchi, Takashi and Watanabe, Kenji and Young, Andrea}, booktitle = {APS March Meeting 2020}, issn = {0003-0503}, location = {Denver, CO, United States}, number = {1}, publisher = {American Physical Society}, title = {{Sublattice resolved spin wave transport through graphene fractional quantum Hall states as a probe of isospin order}}, volume = {65}, year = {2020}, } @inproceedings{10696, abstract = {We experimentally investigate twisted van der Waals heterostructures of monolayer graphene rotated with respect to a bernal stacked graphene bilayer. We report transport measurements for devices with twist angles between 0.9 and 1.4°. The electric field allows efficient tuning of the width, isolation and the topology of the moiré bands in this system. By comparing magnetoresistance measurements to numerical simulations, we develop an understanding of the band structure. Finally, we observe correlated states at half- and quarter-fillings, which arise when narrow moire sublattice band is isolated by energy gaps from dispersive bands. We investigate the effects of in-plane and out-of-plane magnetic field on these states and discuss the implication for their spin- and valley- polarization.}, author = {Polshyn, Hryhoriy and Zhu, Jihang and Kumar, Manish and Taniguchi, Takashi and Watanabe, Kenji and MacDonald, Allan and Young, Andrea}, booktitle = {APS March Meeting 2020}, issn = {0003-0503}, location = {Denver, CO, United States}, number = {1}, publisher = {American Physical Society}, title = {{Correlated states and tunable topological bands in twisted monolayer-bilayer graphene heterostructures}}, volume = {65}, year = {2020}, } @inproceedings{10697, abstract = {We report the observation of a quantized anomalous Hall effect in a moiré heterostructure consisting of twisted bilayer graphene aligned to an encapsulating hBN substrate. The effect occurs at a density of 3 electrons per superlattice unit cell, where we observe magnetic hysteresis and a Hall resistance quantized to within 0.1% of the resistance quantum at temperatures as high as 3K. In this first of 3 talks, I will describe the fabrication procedure for our device as well as basic transport characterization measurements. I will introduce the phenomenology of twisted bilayer graphene and present evidence for hBN alignment as manifested in the hierarchy of symmetry-breaking gaps and anomalous magnetoresistance.}, author = {Zhang, Yuxuan and Serlin, Marec and Tschirhart, Charles and Polshyn, Hryhoriy and Zhu, Jiacheng and Balents, Leon and Huber, Martin E. and Taniguchi, Takashi and Watanabe, Kenji and Young, Andrea}, booktitle = {APS March Meeting 2020}, location = {Denver, CO, United States}, number = {1}, publisher = {American Physical Society}, title = {{Intrinsic quantized anomalous Hall effect in a moiré heterostructure, part I: Device fabrication and transport}}, volume = {65}, year = {2020}, } @inproceedings{10698, abstract = {This is the second of three talks describing the observation and characterization of a ferromagnetic moiré heterostructure based on twisted bilayer graphene aligned to hexagonal boron nitride. I will compare the qualitative and quantitative features of this observed quantum anomalous Hall state to traditional systems engineered from thin film (Bi,Sb)2Te3 topological insulators. In particular, we find that the measured electronic energy gap of ~30K is several times higher than the Curie temperature, consistent with a lack of disorder associated with magnetic dopants. In this system, the quantization arises from spontaneous ferromagnetic polarization into a single spin and valley moiré subband, which is topological despite the lack of spin orbit coupling. I will also discuss the observation of current induced switching, which allows the magnetic state of the heterostructure to be controllably reversed with currents as small as a few nanoamperes.}, author = {Serlin, Marec and Tschirhart, Charles and Polshyn, Hryhoriy and Zhang, Yuxuan and Zhu, Jiacheng and Huber, Martin E. and Balents, Leon and Watanabe, Kenji and Tanaguchi, Takashi and Young, Andrea}, booktitle = {APS March Meeting 2020}, location = {Denver, CO, United States}, number = {1}, publisher = {American Physical Society}, title = {{Intrinsic quantized anomalous Hall effect in a moiré heterostructure, part II: Temperature dependence and current switching}}, volume = {65}, year = {2020}, } @inproceedings{10699, abstract = {This is the third of three talks describing the observation and characterization of a ferromagnetic moiré heterostructure based on twisted bilayer graphene aligned to hexagonal boron nitride. In this segment I will present scanning probe magnetometry data acquired using a nanoSQUID-on-tip microscope, which provides ~150 nm spatial resolution and a field sensitivity of ~10 nT/rtHz. We study the distribution of magnetic domains within the device as a function of density, magnetic field training, and DC current. Our data allow us to constrain the magnitude of the orbital magnetic moment of the electrons in the QAH state. Comparison with simultaneously acquired transport data allows us to precisely correlate single domain dynamics with discrete jumps in the observed anomalous Hall signal.}, author = {Tschirhart, Charles and Serlin, Marec and Polshyn, Hryhoriy and Zhang, Yuxuan and Zhu, Jiacheng and Balents, Leon and Huber, Martin E. and Watanabe, Kenji and Tanaguchi, Takashi and Young, Andrea}, booktitle = {APS March Meeting 2020}, issn = {0003-0503}, location = {Denver, CO, United States}, number = {1}, publisher = {American Physical Society}, title = {{Intrinsic quantized anomalous Hall effect in a moiré heterostructure, part III: Scanning probe magnetometry}}, volume = {65}, year = {2020}, } @article{10701, abstract = {Partially filled Landau levels host competing electronic orders. For example, electron solids may prevail close to integer filling of the Landau levels before giving way to fractional quantum Hall liquids at higher carrier density1,2. Here, we report the observation of an electron solid with non-collinear spin texture in monolayer graphene, consistent with solidification of skyrmions3—topological spin textures characterized by quantized electrical charge4,5. We probe the spin texture of the solids using a modified Corbino geometry that allows ferromagnetic magnons to be launched and detected6,7. We find that magnon transport is highly efficient when one Landau level is filled (ν=1), consistent with quantum Hall ferromagnetic spin polarization. However, even minimal doping immediately quenches the magnon signal while leaving the vanishing low-temperature charge conductivity unchanged. Our results can be understood by the formation of a solid of charged skyrmions near ν=1, whose non-collinear spin texture leads to rapid magnon decay. Data near fractional fillings show evidence of several fractional skyrmion solids, suggesting that graphene hosts a highly tunable landscape of coupled spin and charge orders.}, author = {Zhou, Haoxin and Polshyn, Hryhoriy and Taniguchi, Takashi and Watanabe, Kenji and Young, Andrea F.}, issn = {1745-2481}, journal = {Nature Physics}, number = {2}, pages = {154--158}, publisher = {Springer Nature}, title = {{Skyrmion solids in monolayer graphene}}, doi = {10.1038/s41567-019-0729-8}, volume = {16}, year = {2020}, } @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{9526, abstract = {DNA methylation and histone H1 mediate transcriptional silencing of genes and transposable elements, but how they interact is unclear. In plants and animals with mosaic genomic methylation, functionally mysterious methylation is also common within constitutively active housekeeping genes. Here, we show that H1 is enriched in methylated sequences, including genes, of Arabidopsis thaliana, yet this enrichment is independent of DNA methylation. Loss of H1 disperses heterochromatin, globally alters nucleosome organization, and activates H1-bound genes, but only weakly de-represses transposable elements. However, H1 loss strongly activates transposable elements hypomethylated through mutation of DNA methyltransferase MET1. Hypomethylation of genes also activates antisense transcription, which is modestly enhanced by H1 loss. Our results demonstrate that H1 and DNA methylation jointly maintain transcriptional homeostasis by silencing transposable elements and aberrant intragenic transcripts. Such functionality plausibly explains why DNA methylation, a well-known mutagen, has been maintained within coding sequences of crucial plant and animal genes.}, author = {Choi, Jaemyung and Lyons, David B. and Kim, M. Yvonne and Moore, Jonathan D. and Zilberman, Daniel}, issn = {1097-4164}, journal = {Molecular Cell}, number = {2}, pages = {310--323.e7}, publisher = {Elsevier}, title = {{DNA methylation and histone H1 jointly repress transposable elements and aberrant intragenic transcripts}}, doi = {10.1016/j.molcel.2019.10.011}, volume = {77}, year = {2020}, } @article{9573, abstract = {It is a classical fact that for any ε>0, a random permutation of length n=(1+ε)k2/4 typically contains a monotone subsequence of length k. As a far-reaching generalization, Alon conjectured that a random permutation of this same length n is typically k-universal, meaning that it simultaneously contains every pattern of length k. He also made the simple observation that for n=O(k2logk), a random length-n permutation is typically k-universal. We make the first significant progress towards Alon's conjecture by showing that n=2000k2loglogk suffices.}, author = {He, Xiaoyu and Kwan, Matthew Alan}, issn = {1469-2120}, journal = {Bulletin of the London Mathematical Society}, number = {3}, pages = {515--529}, publisher = {Wiley}, title = {{Universality of random permutations}}, doi = {10.1112/blms.12345}, volume = {52}, year = {2020}, } @article{9576, abstract = {In 1989, Rota made the following conjecture. Given n bases B1,…,Bn in an n-dimensional vector space V⁠, one can always find n disjoint bases of V⁠, each containing exactly one element from each Bi (we call such bases transversal bases). Rota’s basis conjecture remains wide open despite its apparent simplicity and the efforts of many researchers (e.g., the conjecture was recently the subject of the collaborative “Polymath” project). In this paper we prove that one can always find (1/2−o(1))n disjoint transversal bases, improving on the previous best bound of Ω(n/logn)⁠. Our results also apply to the more general setting of matroids.}, author = {Bucić, Matija and Kwan, Matthew Alan and Pokrovskiy, Alexey and Sudakov, Benny}, issn = {1687-0247}, journal = {International Mathematics Research Notices}, number = {21}, pages = {8007--8026}, publisher = {Oxford University Press}, title = {{Halfway to Rota’s basis conjecture}}, doi = {10.1093/imrn/rnaa004}, volume = {2020}, year = {2020}, } @article{9577, abstract = {An n-vertex graph is called C-Ramsey if it has no clique or independent set of size Clogn⁠. All known constructions of Ramsey graphs involve randomness in an essential way, and there is an ongoing line of research towards showing that in fact all Ramsey graphs must obey certain “richness” properties characteristic of random graphs. Motivated by an old problem of Erd̋s and McKay, recently Narayanan, Sahasrabudhe, and Tomon conjectured that for any fixed C, every n-vertex C-Ramsey graph induces subgraphs of Θ(n2) different sizes. In this paper we prove this conjecture.}, author = {Kwan, Matthew Alan and Sudakov, Benny}, issn = {1687-0247}, journal = {International Mathematics Research Notices}, number = {6}, pages = {1621–1638}, publisher = {Oxford University Press}, title = {{Ramsey graphs induce subgraphs of quadratically many sizes}}, doi = {10.1093/imrn/rny064}, volume = {2020}, year = {2020}, } @article{9578, abstract = {How long a monotone path can one always find in any edge-ordering of the complete graph Kn? This appealing question was first asked by Chvátal and Komlós in 1971, and has since attracted the attention of many researchers, inspiring a variety of related problems. The prevailing conjecture is that one can always find a monotone path of linear length, but until now the best known lower bound was n2/3-o(1). In this paper we almost close this gap, proving that any edge-ordering of the complete graph contains a monotone path of length n1-o(1).}, author = {Bucić, Matija and Kwan, Matthew Alan and Pokrovskiy, Alexey and Sudakov, Benny and Tran, Tuan and Wagner, Adam Zsolt}, issn = {1565-8511}, journal = {Israel Journal of Mathematics}, number = {2}, pages = {663--685}, publisher = {Springer}, title = {{Nearly-linear monotone paths in edge-ordered graphs}}, doi = {10.1007/s11856-020-2035-7}, volume = {238}, year = {2020}, } @article{9581, abstract = {We show that for any 𝑛 divisible by 3, almost all order- 𝑛 Steiner triple systems have a perfect matching (also known as a parallel class or resolution class). In fact, we prove a general upper bound on the number of perfect matchings in a Steiner triple system and show that almost all Steiner triple systems essentially attain this maximum. We accomplish this via a general theorem comparing a uniformly random Steiner triple system to the outcome of the triangle removal process, which we hope will be useful for other problems. Our methods can also be adapted to other types of designs; for example, we sketch a proof of the theorem that almost all Latin squares have transversals.}, author = {Kwan, Matthew Alan}, issn = {1460-244X}, journal = {Proceedings of the London Mathematical Society}, number = {6}, pages = {1468--1495}, publisher = {Wiley}, title = {{Almost all Steiner triple systems have perfect matchings}}, doi = {10.1112/plms.12373}, volume = {121}, year = {2020}, } @article{9582, abstract = {The problem of finding dense induced bipartite subgraphs in H-free graphs has a long history, and was posed 30 years ago by Erdős, Faudree, Pach and Spencer. In this paper, we obtain several results in this direction. First we prove that any H-free graph with minimum degree at least d contains an induced bipartite subgraph of minimum degree at least cH log d/log log d, thus nearly confirming one and proving another conjecture of Esperet, Kang and Thomassé. Complementing this result, we further obtain optimal bounds for this problem in the case of dense triangle-free graphs, and we also answer a question of Erdœs, Janson, Łuczak and Spencer.}, author = {Kwan, Matthew Alan and Letzter, Shoham and Sudakov, Benny and Tran, Tuan}, issn = {1439-6912}, journal = {Combinatorica}, number = {2}, pages = {283--305}, publisher = {Springer}, title = {{Dense induced bipartite subgraphs in triangle-free graphs}}, doi = {10.1007/s00493-019-4086-0}, volume = {40}, year = {2020}, } @article{9583, abstract = {We show that for any n divisible by 3, almost all order-n Steiner triple systems admit a decomposition of almost all their triples into disjoint perfect matchings (that is, almost all Steiner triple systems are almost resolvable).}, author = {Ferber, Asaf and Kwan, Matthew Alan}, issn = {2050-5094}, journal = {Forum of Mathematics}, publisher = {Cambridge University Press}, title = {{Almost all Steiner triple systems are almost resolvable}}, doi = {10.1017/fms.2020.29}, volume = {8}, 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}, }