Imperial College London

Networks are an imperfect representation of a dataset, yet often there is little consideration for how these imperfections may affect network evolution and structure.

In this talk, I want to discuss a simple set of generative network models in which the mechanism of network growth is decomposed into two layers. The first layer represents the “observed” network, corresponding to our conventional understanding of a network. Here I want to consider the scenario in which the network you observe is not self-contained, but is driven by a second hidden network, comprised of the same nodes but different edge structure. I will show how a range of different network growth models can be constructed such that the observed and hidden networks can be causally decoupled, coupled only in one direction, or coupled in both directions.

One consequence of such models is the emergence of abrupt transitions in observed network topology – one example results in scale-free degree distributions which are robust up to an arbitrarily long threshold time, but which naturally break down as the network grows larger. I will argue that such examples illustrate why we should be wary of an overreliance on static networks (measured at only one point in time), and will discuss other possible implications for prediction on networks.

Computers can now beat humans at chess and at go. Surely one day they will beat us at proving theorems. But when will it happen, how will it happen, and what should humans be doing in order to make it happen? Furthermore -- do we actually want it to happen? Will they generate incomprehensible proofs, which teach us nothing? Will they find mistakes in the human literature?

I will talk about how I am training undergraduates at Imperial College London to do their problem sheets in a formal proof verification system, and how this gamifies mathematics. I will talk about mistakes in the modern pure mathematics literature, and ask what the point of modern pure mathematics is.

We discuss asymptotics of Toeplitz determinants with Fisher--Hartwig singularities, and give an overview of past and more recent results.
Applications include the study of asymptotics of certain statistics of the characteristic polynomial of the Circular Unitary Ensemble (CUE) of random matrices. In particular recent results in the study of Toeplitz determinants allow for a proof of a conjecture by Fyodorov and Keating on moments of averages of the characteristic polynomial of the CUE.
 

Many complex systems can be represented as networks. However, it is often not possible or even desirable to observe the entire network structure. For example, in social networks, it is often difficult to obtain samples of large networks due to commercial sensitivity or privacy concerns relating to the data. However, it may be possible to provide a coarse grained picture of the graph given knowledge of the distribution of different demographics (e.g age, income, location, etc…) in a population and their propensities for forming ties between each other.

I will explore the degree to which it is possible to influence Ising systems, which are commonly used to model social influence, on unobserved graphs. Using both synthetic networks (stochastic blockmodels) and case studies of real world social networks, I will demonstrate how simple models which rely only on a coarse grained description of the system or knowledge of only the underlying external fields can perform comparably to more expensive optimization algorithms.

I will explore the geometrical structure of Higgs branches of quantum field theories with 8 supercharges in 3, 4, 5 and 6 dimensions. They are hyperkahler singularities, and as such they can be described by a Hasse diagram built from a family of elementary transitions. This corresponds physically to the partial Higgs mechanism. Using brane systems and recently introduced notions of magnetic quivers and quiver subtraction, we formalise the rules to obtain the Hasse diagrams.

Complexity Science aims to understand what is that makes some systems to be "more than the sum of their parts". A natural first step to address this issue is to study networks of pairwise interactions, which have been done with great success in many disciplines -- to the extend that many people today identify Complexity Science with network analysis. In contrast, multivariate complexity provides a vast and mostly unexplored territory. As a matter of fact, the "modes of interdependency" that can exist between three or more variables are often nontrivial, poorly understood and, yet, are paramount for our understanding of complex systems in general, and emergence in particular. 
In this talk we present an information-theoretic framework to analyse high-order correlations, i.e. statistical dependencies that exist between groups of variables that cannot be reduced to pairwise interactions. Following the spirit of information theory, our approach is data-driven and model-agnostic, being applicable to discrete, continuous, and categorical data. We review the evolution of related ideas in the context of theoretical neuroscience, and discuss the most prominent extensions of information-theoretic metrics to multivariate settings. Then, we introduce the O-information, a novel metric that quantify various structural (i.e. synchronous) high-order effects. Finally, we provide a critical discussion on the framework of Integrated Information Theory (IIT), which suggests an approach to extend the analysis to dynamical settings. To illustrate the presented methods, we show how the analysis of high-order correlations can reveal critical structures in various scenarios, including cellular automata, Baroque music scores, and various EEG datasets.

References:
[1] F. Rosas, P.A. Mediano, M. Gastpar and H.J. Jensen, ``Quantifying High-order Interdependencies via Multivariate Extensions of the Mutual Information'', submitted to PRE, under review.
https://arxiv.org/abs/1902.11239
[2] F. Rosas, P.A. Mediano, M. Ugarte and H.J. Jensen, ``An information-theoretic approach to self-organisation: Emergence of complex interdependencies in coupled dynamical systems'', in Entropy, vol. 20 no. 10: 793, pp.1-25, Sept. 2018.
https://www.mdpi.com/1099-4300/20/10/793

 

This work aims at developing new approach for parameterizing mesoscale eddy effects for use in non-eddy-resolving ocean circulation models. These effects are often modelled as some diffusion process or a stochastic forcing, and the proposed approach is implicitly related to the latter category. The idea is to approximate transient eddy flux divergence in a simple way, to find its actual dynamical footprints by solving a simplified but dynamically relevant problem, and to relate the ensemble of footprints to the large-scale flow properties.

In the first part of the talk, we study risk-sharing equilibria where heterogenous agents trade subject to quadratic transaction costs. The corresponding equilibrium asset prices and trading strategies are characterised by a system of nonlinear, fully-coupled forward-backward stochastic differential equations. We show that a unique solution generally exists provided that the agents’ preferences are sufficiently similar. In a benchmark specification, the illiquidity discounts and liquidity premia observed empirically correspond to a positive relationship between transaction costs and volatility.
In the second part of the talk, we discuss how the model can be calibrated to time series of prices and the corresponding trading volume, and explain how extensions of the model with general transaction costs, for example, can be solved numerically using the deep learning approach of Han, Jentzen, and E (2018).
 (Based on joint works with Martin Herdegen and Dylan Possamai, as well as with Lukas Gonon and Xiaofei Shi)

In the past decade, deep learning methods have achieved unprecedented performance on a broad range of problems in various fields from computer vision to speech recognition. So far research has mainly focused on developing deep learning methods for Euclidean-structured data. However, many important applications have to deal with non-Euclidean structured data, such as graphs and manifolds. Such data are becoming increasingly important in computer graphics and 3D vision, sensor networks, drug design, biomedicine, high energy physics, recommendation systems, and social media analysis. The adoption of deep learning in these fields has been lagging behind until recently, primarily since the non-Euclidean nature of objects dealt with makes the very definition of basic operations used in deep networks rather elusive. In this talk, I will introduce the emerging field of geometric deep learning on graphs and manifolds, overview existing solutions and outline the key difficulties and future research directions. As examples of applications, I will show problems from the domains of computer vision, graphics, high-energy physics, and fake news detection.