TY - JOUR AB - For a general class of large non-Hermitian random block matrices X we prove that there are no eigenvalues away from a deterministic set with very high probability. This set is obtained from the Dyson equation of the Hermitization of X as the self-consistent approximation of the pseudospectrum. We demonstrate that the analysis of the matrix Dyson equation from (Probab. Theory Related Fields (2018)) offers a unified treatment of many structured matrix ensembles. AU - Alt, Johannes AU - Erdös, László AU - Krüger, Torben H AU - Nemish, Yuriy ID - 6240 IS - 2 JF - Annales de l'institut Henri Poincare SN - 0246-0203 TI - Location of the spectrum of Kronecker random matrices VL - 55 ER - TY - JOUR AB - Long non-coding (lnc) RNAs are numerous and found throughout the mammalian genome, and many are thought to be involved in the regulation of gene expression. However, the majority remain relatively uncharacterised and of uncertain function making the use of model systems to uncover their mode of action valuable. Imprinted lncRNAs target and recruit epigenetic silencing factors to a cluster of imprinted genes on the same chromosome, making them one of the best characterized lncRNAs for silencing distant genes in cis. In this study we examined silencing of the distant imprinted gene Slc22a3 by the lncRNA Airn in the Igf2r imprinted cluster in mouse. Previously we proposed that imprinted lncRNAs may silence distant imprinted genes by disrupting promoter-enhancer interactions by being transcribed through the enhancer, which we called the enhancer interference hypothesis. Here we tested this hypothesis by first using allele-specific chromosome conformation capture (3C) to detect interactions between the Slc22a3 promoter and the locus of the Airn lncRNA that silences it on the paternal chromosome. In agreement with the model, we found interactions enriched on the maternal allele across the entire Airn gene consistent with multiple enhancer-promoter interactions. Therefore, to test the enhancer interference hypothesis we devised an approach to delete the entire Airn gene. However, the deletion showed that there are no essential enhancers for Slc22a2, Pde10a and Slc22a3 within the Airn gene, strongly indicating that the Airn RNA rather than its transcription is responsible for silencing distant imprinted genes. Furthermore, we found that silent imprinted genes were covered with large blocks of H3K27me3 on the repressed paternal allele. Therefore we propose an alternative hypothesis whereby the chromosome interactions may initially guide the lncRNA to target imprinted promoters and recruit repressive chromatin, and that these interactions are lost once silencing is established. AU - Andergassen, Daniel AU - Muckenhuber, Markus AU - Bammer, Philipp C. AU - Kulinski, Tomasz M. AU - Theussl, Hans-Christian AU - Shimizu, Takahiko AU - Penninger, Josef M. AU - Pauler, Florian AU - Hudson, Quanah J. ID - 7399 IS - 7 JF - PLoS Genetics SN - 1553-7404 TI - The Airn lncRNA does not require any DNA elements within its locus to silence distant imprinted genes VL - 15 ER - TY - JOUR AB - Origin and functions of intermittent transitions among sleep stages, including short awakenings and arousals, constitute a challenge to the current homeostatic framework for sleep regulation, focusing on factors modulating sleep over large time scales. Here we propose that the complex micro-architecture characterizing the sleep-wake cycle results from an underlying non-equilibrium critical dynamics, bridging collective behaviors across spatio-temporal scales. We investigate θ and δ wave dynamics in control rats and in rats with lesions of sleep-promoting neurons in the parafacial zone. We demonstrate that intermittent bursts in θ and δ rhythms exhibit a complex temporal organization, with long-range power-law correlations and a robust duality of power law (θ-bursts, active phase) and exponential-like (δ-bursts, quiescent phase) duration distributions, typical features of non-equilibrium systems self-organizing at criticality. Crucially, such temporal organization relates to anti-correlated coupling between θ- and δ-bursts, and is independent of the dominant physiologic state and lesions, a solid indication of a basic principle in sleep dynamics. AU - Wang, Jilin W. J. L. AU - Lombardi, Fabrizio AU - Zhang, Xiyun AU - Anaclet, Christelle AU - Ivanov, Plamen Ch. ID - 7103 IS - 11 JF - PLoS Computational Biology SN - 1553-7358 TI - Non-equilibrium critical dynamics of bursts in θ and δ rhythms as fundamental characteristic of sleep and wake micro-architecture VL - 15 ER - TY - CONF AB - Knowledge distillation, i.e. one classifier being trained on the outputs of another classifier, is an empirically very successful technique for knowledge transfer between classifiers. It has even been observed that classifiers learn much faster and more reliably if trained with the outputs of another classifier as soft labels, instead of from ground truth data. So far, however, there is no satisfactory theoretical explanation of this phenomenon. In this work, we provide the first insights into the working mechanisms of distillation by studying the special case of linear and deep linear classifiers. Specifically, we prove a generalization bound that establishes fast convergence of the expected risk of a distillation-trained linear classifier. From the bound and its proof we extract three keyfactors that determine the success of distillation: data geometry – geometric properties of the datadistribution, in particular class separation, has an immediate influence on the convergence speed of the risk; optimization bias– gradient descentoptimization finds a very favorable minimum of the distillation objective; and strong monotonicity– the expected risk of the student classifier always decreases when the size of the training set grows. AU - Bui Thi Mai, Phuong AU - Lampert, Christoph ID - 6569 T2 - Proceedings of the 36th International Conference on Machine Learning TI - Towards understanding knowledge distillation VL - 97 ER - TY - CONF AB - Modern machine learning methods often require more data for training than a single expert can provide. Therefore, it has become a standard procedure to collect data from external sources, e.g. via crowdsourcing. Unfortunately, the quality of these sources is not always guaranteed. As additional complications, the data might be stored in a distributed way, or might even have to remain private. In this work, we address the question of how to learn robustly in such scenarios. Studying the problem through the lens of statistical learning theory, we derive a procedure that allows for learning from all available sources, yet automatically suppresses irrelevant or corrupted data. We show by extensive experiments that our method provides significant improvements over alternative approaches from robust statistics and distributed optimization. AU - Konstantinov, Nikola H AU - Lampert, Christoph ID - 6590 T2 - Proceedings of the 36th International Conference on Machine Learning TI - Robust learning from untrusted sources VL - 97 ER -