Giuseppe Ughi

MSc in Mathematical Modelling and Scientific Computing (University of Oxford), BSc in Mathematical Engineering (Politecnico di Milano)

Status

Postdoctoral Research Associate

Contact form

+44 1865 270508

Research groups

Address
Mathematical Institute
University of Oxford
Andrew Wiles Building
Radcliffe Observatory Quarter
Woodstock Road
Oxford
OX2 6GG

Recent publications

Mutual information of neural network initialisations: mean field approximations

Tanner, J Ughi, G Proceedings of the IEEE International Symposium on Information Theory 813-818 (01 Sep 2021)

An empirical study of derivative-free-optimization algorithms for targeted black-box attacks in deep neural networks

Ughi, G Abrol, V Tanner, J Optimization and Engineering (30 Jun 2021)

Research interests

The main focus of my research is on Theories of Deep Learning with the aim of applying them to the automatisation of the appraisal of real-estates with RE5Q.

Initially, I focused on the application of traditional derivative-free optimisation algorithms to generate adversarial examples for deep nets. This lead to a comprehensive review of how different optimisation techniques compare in this task.

Currently, I am focusing on the use of Information Theory to develop initialisations of neural networks that allow an optimal transmission of information through the layers.

Major / recent publications

Invariant Risk Minimisation for Cross-Organism Inference: Substituting Mouse Data for Human Data in Human Risk Factor Discovery

O O'Donoghue, P Duckworth, G Ughi, L Scheibenreif, K Khezeli, A Hoarfrost, S Budd, P Foley, N Chia, J Kalantari, G Mackintosh, F Soboczenski, L Sanders

NeurIPS 2021 Workshop Machine Learning for Health (2021)

An empirical study of derivative-free-optimization algorithms for targeted black-box attacks in deep neural networks

G Ughi, V Abrol, J Tanner

Optimization and Engineering, Springer (2021)

Mutual Information of Neural Network Initialisations: Mean Field Approximations

J Tanner, G Ughi

2021 IEEE International Symposium on Information Theory (ISIT 2021)

A Model-Based Derivative-Free Approach to Black-Box Adversarial Examples: BOBYQA

G Ughi, V Abrol, J Tanner

NeurIPS 2019 Workshop: Beyond First Order Methods in ML (2019)

Further details

Post-Doctoral Research Associate within the CIMDA-Oxford collaboration.

Teaching

Tutor for:

C.6.5 Theories of Deep Learning (MT20)
B.5.6 Nonlinear Systems (HT20)
C.6.5 Theories of Deep Learning (MT19)

Teaching Assistant for:

C.7.7 Random Matrix Theory (HT21)
Oxford Internet Institute - Machine Learning (MT20)
C.6.5 Theories of Deep Learning (HT19)
B.8.4 Information Theory (MT18)