CoSy Seminars Spring 2018

13 February
Speaker: Douglas Hofstadter, Center for Research on Concepts and Cognition, Indiana University
Title: Life and Death in a Family of Meta-Fibonacci Recursions
Place: Å4003
Time: 12:15 -- 13:00

Abstract: See attached file

27 February
Speaker: Elisenda Feliu, Department of Mathematical Sciences, University of Copenhagen
Title: An algebraic approach to biochemical reaction networks
Place: Å4003
Time: 12:15 -- 13:00

Abstract: Under the law of mass-action, the concentrations of the species of a chemical reaction network are often modelled by means of a system of polynomial ordinary differential equations. The polynomial structure of the system allows us to use techniques that come from algebra and graph theory to study properties of the system, mainly at steady state. This is the basis of the so-called Chemical Reaction Network Theory. In this talk I will start by presenting the formalism of the approach and discuss the questions we would like to address. I will proceed to present some of the results we have recently obtained, mainly focusing on the number of steady states of the system.

6 Mars
Speaker: Christina Kuttler, Zentrum Mathematik, Technische Universität München
Title: Feedbacks in bacterial systems using intraspecies communication
Place: Å4003
Time: 12:15 -- 13:00

Abstract: Quorum sensing (QS) is a mechanism which enables
many bacterial species to coordinate their behaviour via release and
uptake of signalling molecules.

More and more bacterial species using such a
mechanisms are known, including many details about the underlying
gene regulation processes. QS may control very different bacterial
processes, like pathogeneity, production of public and private goods,
but also to some extent resistance against „stressors“. 

Within the gene regulation systems, but also on
the population level, there are several positive and negative
feedback loops interacting in a complex, highly nonlinear way.
Mathematical modelling helps to understand their influences on the
dynamics of such systems. We introduce mathematical modelling
approaches in the form of ODEs, DDEs and PDEs. 

One goal of a better understanding is the
potential usage of these processes for a better control of bacterial
populations and their behaviour. 

13 Mars
Speaker: David Garcia, Complexity Science Hub Vienna and Medical University of Vienna
Title: Collective emotions and social resilience in the digital traces after a terrorist attack
Place: Å2004
Time: 12:15 -- 13:00

Abstract: After collective traumas like natural disasters and terrorist attacks, members of concerned communities experience intense emotions and talk profusely about them. These verbal exchanges resemble emotional venting and seem devoid of social functions. However, Durkheim’s theory of emotional effervescence postulates that these collective emotions fulfill major social functions, generating social identity, reinforcing shared beliefs, and leading to higher solidarity. We present the first large-scale test of this theory through the longitudinal analysis of digital traces captured in Twitter after the Paris terrorist attacks of November, 2015. Examining the temporal evolution of these collective emotional responses, we observe them to last considerably longer than emotions in isolation. Collective emotional expression is followed by a marked increase in the use of lexical indicators related to social resilience, in particular social processes, prosocial behavior, and shared values. In addition, we show that individuals who participated to a higher degree in the collective emotion also evidenced a superior use of terms associated to social resilience in the months after, though they did not evidence this trend in the months before the attacks. Together, our findings support the  conclusion that collective traumas can activate emotion sharing feedback loops in the concerned community, as described by Durkheim. Our results support the existence of social resilience effects following the collective emotions elicited by a terrorist attack.

20 Mars
Speaker: Fred Hamprecht, Interdisciplinary Center for Scientific Computing (IWR) and
Department of Physics and Astronomy, Heidelberg University
Title: The quest for the wiring diagram of the brain - Where computer vision, deep neural networks and combinatorial
optimization meet
Place: Å4003
Time: 12:15 -- 13:00

Abstract: Understanding the brain is an old and yet-unsolved problem. To
understand the workings of a neural circuit, it is possibly required to
know its structure, and almost surely necessary to know its connectivity.
After great progress in electron microscopy, several labs worldwide are
milling away at animal brains and generating what will amount to
petabytes of high-quality data. The resulting images are good enough for
human tracers to consistently follow at least the majority of neural
processes; unfortunately, humans would take thousands of years to
complete the task for even the smallest mammalian brain.
So the quest is on for computer vision algorithms to do the same
automatically and reliably. The current state of the art pipelines recur
to deep neural networks and combinatorial graph partitioning problems.
The former are notoriously ill understood, the latter still expensive to
solve at scale. In this talk, I will sketch the problem, a state of the art approach
(which does not quite achieve human accuracy yet), and I will lay out
some of the open problems in the field.

24 April
Speaker: Ekaterina Fetisova, Department of Mathematics, Stockholm University
Title: Towards a flexible statistical modelling by latent factors for evaluation of simulated responses to climate forcings
Place: Å4003
Time: 12:15 -- 13:00

Abstract: In my talk, I will present a new flexible statistical framework developed for evaluation of temperature data, generated by sophisticated climate models, against observational data. The framework includes several so-called Structural Equation Models (SEM) with latent variables, subsuming also such models as Measurement Error (ME) models and Confirmatory Factor Analysis (CFA) models. Evaluation of climate model simulations is accomplished by assessing whether latent temperature responses to different climate forcings (such as solar, orbital, volcanic etc.) are correctly represented in climate model simulations, compared to the corresponding true temperature responses embedded in observational data. In addition, the framework allows researchers to investigate the underlying latent structure of temperature data and to make statistical inferences about influences of the forcings on the temperature. The flexibility of the framework is reflected in its ability to take into account the number of forcings, their climate-relevant properties, and our substantive knowledge about possible (causal) relationships between temperature responses to various forcings. The performance of some ME-, CFA- and SEM-models is evaluated and compared in a pseudo-proxy experiment, where observational data are replaced by temperature data from a selected climate model simulations. This replacement ensures that latent temperature responses are indeed correctly represented in the climate model simulations under study. Nevertheless, the underlying structure remains unknown. The results of the experiment indicated that the underlying latent structure of the temperature data under consideration is too complicated to be described by ME models. Depending on the region under consideration, either a CFA- or SEM-model is required to provide an acceptable and climatologically defensible description of the underlying latent structure of the temperature data.


8 May
Speaker: Mattias Villani, Department of Computer and Information Science, Linköping University
Title: The Block-Poisson Estimator for Optimally Tuned Exact Subsampling MCMC
Place: Å4003
Time: 12:15 -- 13:00

Abstract: Markov Chain Monte Carlo (MCMC) algorithms are the workhorse for Bayesian inference. In large datasets with many observations there is a recent trend to use the pseudo-marginal MCMC framework where the likelihood is estimated from a subset of the data in each MCMC iteration. This still targets the correct posterior distribution if the estimator is unbiased, and can be substantially faster than regular MCMC. In this talk, I will present some of our recent work on subsampling MCMC based on a new unbiased block-Poisson estimator of the likelihood. We derive some key properties of this estimator which are used to find guidelines for selecting the optimal subsample size. The guidelines apply to any pseudo-marginal algorithm with the likelihood estimated by the block-Poisson estimator, including the important class of doubly intractable problems.

22 May

Speaker: Juan Fernández-Gracia, Institute for Cross-Disciplinary Physics and Complex Systems, University of the Balearic Islands and Spanish Research Council (UIB-CSIC)
Title: Electoral data analysis and modeling
Place: Å4003
Time: 12:15 -- 13:00

Abstract: Election data let us take a look at human behavior. Besides deciding our future, these data are very useful for the study of collective social phenomena, as they are spatially extended and we have data for several decades in many countries, which allows for spatial and temporal studies. Besides, typically these data are publicly available, at least in modern democracies.

In this talk I will review two different ways of looking at electoral processes that are complementary and sometimes lack a connection. On the one side there is a data analysis based approach which aims at detecting patterns in the data (or the lack of them), while on the other side there is the more theoretical model based approach that tries to disentangle the mechanisms at play in collective decision making.

11 June

Speaker: Lieven De Lathauwer, KU Leuven, Belgium
Title: An introduction to tensor methods
Place: Å4003
Time: 12:15 -- 13:00

This talk is meant as a low-level introduction to some of the remarkable features of higher-order tensors and their relevance for applications. First we pay attention to the basic Canonical Polyadic Decomposition and the Tucker decomposition. The uniqueness properties of CPD make it a powerful tool for signal separation and data analysis. Block term decompositions allow us to retrieve components that are more general and possibly more realistic than rank-1 terms. Multilinear singular value decomposition and low multilinear rank approximation are key in multilinear extensions of subspace techniques. Coupled decompositions express complicated tasks as combinations of pieces that can be handled. Tensor trains and hierarchical Tucker decompositions allow one to break the curse of dimensionality in a numerically reliable manner and show promise for big data analysis in combination with compressed sensing. Time permitting, we will touch on some computational aspects and how they have been dealt with in Tensorlab (