Jose Such (King's College)

Title: Multiparty Privacy and AI Security & Privacy

Abstract:
In this talk, I will provide an overview of two problems I have been contributing to in the past years: i) Multiparty Privacy; ii) AI Security & Privacy. Regarding i), privacy is not just about what an individual user discloses about herself, it also involves what others (e.g. her friends) may disclose about her. Multiparty privacy is concerned with information pertaining to several individuals and the conflicts that arise when the privacy preferences of these individuals differ. In particular, I will go through the research we have been doing on multiparty privacy in a social media context and the future challenges in general in this field. Regarding ii) I will be talking about the work we have been doing at the intersection of HCI and AI on security and privacy in AI-based systems, including privacy threats in AI-based systems, users’ perceptions of security and privacy in AI-based systems, socio-technical approaches to tackle AI-based discrimination against particular type of users (e.g. based on gender/ethnicity), and symbolic AI approaches that can help towards addressing security, privacy and transparency issues in AI-based systems.

Carl Henrik Ek (Bristol)

Title: Modulated Bayesian Optimisation

Abstract:
Bayesian optimization (BO) methods often rely on the assumption that the objective function is well-behaved, but in practice, this is seldom true. Common approaches, which try to model the objective as precisely as possible, often fail to make progress by spending too many evaluations modelling irrelevant details. In this talk, I will introduce a set of surrogate models that aims focuses on the details that are informative for search while ignoring detrimental structure that is challenging to model from a few observations. We demonstrate that surrogate models with appropriate noise distributions can absorb challenging structures in the objective function by treating them as irreducible uncertainty.

Tugkan Batu (LSE)

Title: Testing Discrete Distributions -- a CS Retrospective

Abstract: Making inferences about the distributional properties of data is arguably one of the most fundamental classes of data analysis tasks. The study of such inferences has been a scientific endeavour for a long time, typically, by statistics and information theory. In the last two decades, testing problems for distributions, especially, on discrete domains have been studied also in the theoretical computer science community, often from a slightly different perspective. The primary objective in this line of research is characterising the number of samples needed, in terms of the domain size or some function of the distributions involved, to perform a given inference task reliably. In this talk, I will present an overview of this very active field of research, describing some basic models and problems, landmark results, and future directions.

Ata Kaban (Birmingham)

Title: Uncovering structure with randomness for learning in high dimensions

Abstract:
Random projection (RP) is a simple, computationally efficient linear
dimensionality reduction technique with interesting theoretical
properties for learning from data. For instance, it can regularise
ill-conditioned covariance matrices, and it preserves dot products
between vectors and their signs to an extent that depends on their
cosine similarity. It can speed up computations when dealing with high
dimensional data sets, while retaining control of generalisation
performance. Furthermore, in this talk we discuss how RP can also be
used as an analytic tool to better understand why learning from high
dimensional data is not always as difficult as it looks. We present
generalisation bounds where the use of RP highlights how learning can
take advantage of the presence of certain benign geometric structures
in the problem, and yields conditions that reduce
dimensional-dependence of error-guarantees in settings where such
dependence is known to be essential in general.

Samson Abramsky (Oxford)

Title: Non-classicality, quantum resources and quantum advantage (slides)

Abstract:
A key objective in the field of quantum information and computation is to understand the advantage which can be gained in information processing tasks by the use of quantum resources. While a range of examples has been studied, to date a systematic understanding of quantum advantage is lacking.

Our focus here is on quantum resources which take the form of non-local, or more generally \emph{contextual}, correlations. Contextuality is one of the key signatures of non-classicality in quantum mechanics and has been shown to be a necessary ingredient for quantum advantage in a range of information processing tasks. We will describe a notion of simulation between quantum resources, and more generally between resources described in terms of contextual correlations, in the ``sheaf-theoretic'' framework for contextuality (Abramsky-Brandenburger).
The notion of simulation is expressed as a morphism of empirical models, in a form which allows the behaviour of one set of correlations to be simulated in terms of another using classical processing and shared randomization.

The talk will be essentially self-contained, introducing the various notions we will discuss.

Jakob Nikolas Kather (University Hospital Aachen)

Title: Deep Learning for Precision Oncology

Abstract:
Precision oncology requires molecular and genetic testing of tumour tissue. For many tests, widespread implementation into clinical practise is limited because these biomarkers are costly, require significant expertise and are limited by tissue availability. However, virtually every cancer patient gets a biopsy as part of the diagnostic workup and this tissue is routinely stained with hematoxylin and eosin (HE). We have developed a deep learning-based technology to predict molecular features, prognosis and markers for treatment response directly from routine histology. This talk will summarize the state of the art in deep learning histopathology, demonstrate emerging use cases and discuss the clinical implications of deep learning-based molecular testing of solid tumours.

Edith Elkind (Oxford)

Title: Fair Division of a Graph

Abstract: We consider fair allocation of indivisible items under an additional constraint: there is an undirected graph describing the relationship between the items, and each agent's share must form a connected subgraph of this graph. This framework captures, e.g., fair allocation of land plots, where the graph describes the accessibility relation among the plots. We focus on agents that have additive utilities for the items, and consider several common fair division solution concepts, such as proportionality, envy-freeness and maximin share guarantee.In particular, we prove that for acyclic graphs a maximin share allocation always exists and can be computed efficiently. We also discuss a continuous variant of this problem, where agents need to divide a collection of interconnected "cakes" and each agent's piece has to be connected, or, more generally, have a small number of connected components.

Based on joint work with Sylvain Bouveret, Katarina Cehlarova, Ayumi Igarashi, Dominik Peters, Xiaohui Bei and Warut Suksompong

Mario Berta (Imperial College London)

Title: Quantum Technologies for Cryptography

Abstract:
The rise of quantum technologies opens exciting possibilities for performing cryptographic tasks, but it also poses new threats. On the one hand, quantum mechanical devices are the basis of a variety of novel cryptographic protocols like key distribution or secure identification, which offer a significant advantage over standard routines. On the other hand, quantum computers will render some existing cryptographic schemes insecure. The interplay between these two aspects is best understood by analysing the power of adversaries which have access to a quantum computer - quantum adversaries. In my talk, I will discuss methods to analyse the power of quantum adversaries which shine light on both aspects.