L2

Melanie Weber 

Title: Geometric Methods for Machine Learning and Optimization

Abstract: A key challenge in machine learning and optimization is the identification of geometric structure in high-dimensional data. Such structural understanding is of great value for the design of efficient algorithms and for developing fundamental guarantees for their performance. Motivated by the observation that many applications involve non-Euclidean data, such as graphs, strings, or matrices, we discuss how Riemannian geometry can be exploited in Machine Learning and Optimization. First, we consider the task of learning a classifier in hyperbolic space. Such spaces have received a surge of interest for representing large-scale, hierarchical data, since they achieve better representation accuracy with fewer dimensions. Secondly, we consider the problem of optimizing a function on a Riemannian manifold. Specifically, we will consider classes of optimization problems where exploiting Riemannian geometry can deliver algorithms that are computationally superior to standard (Euclidean) approaches.

Francesca Panero

Title: A general overview of the different projects explored during my DPhil in Statistics.

Abstract: In the first half of the talk, I will present my work on statistical models for complex networks. I will propose a model to describe sparse spatial random graph underpinned by the Bayesian nonparametric theory and asymptotic properties of a more general class of these models, regarding sparsity, degree distribution and clustering coefficients.

The second half will be devoted to the statistical quantification of the risk of disclosure, a quantity used to evaluate the level of privacy that can be achieved by publishing a microdata file without modifications. I propose two ways to estimate the risk of disclosure, using both frequentist and Bayes nonparametric statistics.

*Note the different room location (L2) to usual Fridays@4 sessions*

This week is Mental Health Awareness Week. To mark this, Rebecca Reed from Siendo will deliver a session on mental health and wellbeing. The session will cover the following things: 

- The importance of finding a balance with achievement and managing stress and pressure.
- Coping mechanisms work with stresses at work in a positive way (not seeing all stress as bad).
- The difficulties faced in the HE environment, such as the uncertainty felt within jobs and research, combined with the high expectations and workload. 

Holographic correlation functions are under good analytic control when none of the single trace operators live in long multiplets. This is famously the case for SCFTs with sixteen supercharges but it is also possible to construct examples with eight supercharges by exploiting space filling branes in AdS. In particular, one can study 4d N=2 theories which are related to each other by an S-fold in much the same way that N=3 theories are related to N=4 Super Yang-Mills. I will describe how modern methods provide a window into their correlation functions with an emphasis on anomalous dimensions. To compare the different S-folds we will need to go to one loop, and to go to one loop we will need to account for operator mixing. This provides an example of resolving degeneracy by resolving degeneracy.

I will review aspects of the theory of symmetry-protected topological phases, focusing on the case of one-dimensional quantum chains. Important concepts include the bulk-boundary correspondence, with bulk topological invariants leading to interesting boundary phenomena. I will discuss topological invariants and associated boundary phenomena in the case that the system is gapless and described at low energies by a conformal field theory. Based on work with Ruben Verresen, Ryan Thorngren and Frank Pollmann.

Topological data analysis is starting to establish itself as a powerful and effective framework in machine learning , supporting the analysis of neural networks, but also driving the development of novel algorithms that incorporate topological characteristics. As a problem class, graph representation learning is of particular interest here, since graphs are inherently amenable to a topological description in terms of their connected components and cycles. This talk will provide
an overview of how to address graph learning tasks using machine learning techniques, with a specific focus on how to make such techniques 'topology-aware.' We will discuss how to learn filtrations for graphs and how to incorporate topological information into modern graph neural networks, resulting in provably more expressive algorithms. This talk aims to be accessible to an audience of TDA enthusiasts; prior knowledge of machine learning is helpful but not required.

Adaptive agents through active inference: The main fields of research that are used to model and realise adaptive agents are optimal control, reinforcement learning and active inference. Active inference is a probabilistic description of adaptive agents that is relatively less known to mathematicians, as it originated from neuroscience in the last decade. This talk presents the mathematical underpinnings of active inference, starting from fundamental considerations about agents that maintain their structural integrity in the face of environmental perturbations. Through this, we derive a probability distribution over actions, that describes decision-making under uncertainty in adaptive agents . Interestingly, this distribution has an interesting information geometric structure, combining, for instance, drives for exploration and exploitation, which may yield a principled answer to the exploration-exploitation trade-off. Preserving this geometric structure enables to realise adaptive agents in practice. We illustrate their behaviour with simulation examples and empirical comparisons with reinforcement learning.

Scaling Limits of Random Graphs: The scaling limit of directed random graphs remains relatively unexplored compared to their undirected counterparts. In contrast, many real-world networks, such as links on the world wide web, financial transactions and “follows” on Twitter, are inherently directed. Previous work by Goldschmidt and Stephenson established the scaling limit for the strongly connected components (SCCs) of the Erdős -- Rényi model in the critical window when appropriately rescaled. In this talk, we present a result showing the SCCs of another class of critical random directed graphs will converge when rescaled to the same limit. Central to the proof is an exploration of the directed graph and subsequent encodings of the exploration as real valued random processes. We aim to present this exploration algorithm and other key components of the proof.

From Mathematics to Data Science and Back: We give an overview of the interaction between rough path theory and data science at the current time.
 

 I will review some well-established relationships between four manifolds and vertex algebras that can be deduced from studying the M5-brane worldvolume theory, and outline some of the corresponding results in mathematics which have been understood so far. I will then describe a proposal of Gaiotto-Rapcak to generalize these ideas to the setting of multiple M5 branes wrapping divisors in toric Calabi-Yau threefolds, and explain work in progress on understanding the mathematical implications of this proposal as a complex network of relationships between the enumerative geometry of sheaves on threefolds and the representation theory of affine Lie algebras.

 In this talk I will discuss an algorithm to piecewise dualise linear quivers into their mirror duals. This applies to the 3d N=4 version of mirror symmetry as well as its recently introduced 4d counterpart, which I will review. The algorithm uses two basic duality moves, which mimic the local S-duality of the 5-branes in the brane set-up of the 3d theories, and the properties of the S-wall. The S-wall is known to correspond to the N=4 T[SU(N)] theory in 3d and I will argue that its 4d avatar corresponds to an N=1 theory called E[USp(2N)], which flows to T[SU(N)] in a suitable 3d limit. All the basic duality moves and S-wall properties needed in the algorithm are derived in terms of some more fundamental Seiberg-like duality, which is the Intriligator--Pouliot duality in 4d and the Aharony duality in 3d.