My interests lie in Human-Centered AI - keenly influenced by my passion for Philosophy, Psychology and Sociology. Currently working as a Doctoral Researcher for MyWorld in the Creative-AI industries.

For my MEng I studied at the University of Southampton and wrote my dissertation on philosophical frameworks by relating classification techniques such as PCA and LDA to philosophical and sociological concepts; supervised by Mark Weal.

I am currently pursuing a PhD in Intelligent Cinematography (Y1/3). My project is focused on intelligent 3D reconstruction of virtual stage/volume, film production, environments for use in previs. My goal is to provide this as an Unreal 5 plugin/configurable workspace for 3D set-design, advanced 3D stroyboarding, simulating shots, and so on. The research goal is to reconstruct a physcial workspace from one video sweep of the environment and recognise equiptment and structures to correct and segment, ultimately providing highquality configurable assets. The project is supervised by Dave Bull and Pui Anantrasirichai and funded by My World. Call for collaboration: I am looking for studios willing to provide access to virtual stages and related equiptment for 3D capture (not for use of a stage). This could either be a singular capture of the stage space (containing equitpment) or capture of individual/grouped equiptment. If you're interested in collaborating, shoot me an email.

Reasonably and Reliably Designing Human-tasked Robots through Reignforcement Learning

Literature Review - Jan-June 2022

I have evaluated a dozen of papers which explore designing human-centric AI by training a network using numerous reinforcement learning methods. Many of these papers have been taken from AIES 2020/21 conference. (On going)

Solving a Maze blind using a Duelling Deep Q-Network

Model Implementation and Extension - Jan-June 2022

As a result of the progression made with deep Q-learning, we assess methods of optimisation which will allow our actor, who is globally-blind (can only observe their direct surroundings), to solve a large maze. We have inspected the use of combined experience replay (A Deeper Look at Experience Replay, Zhang and Sutton, 2017) for our transition buffer, as well as compared the greedy episilon and Boltzmann eplxoration methods. Repository for this project can be found here.

The Importance of Group-Size Preferences for the Evolution of Cooperation Under the Conditions of Individual Selection

Literature Review and Model Extension - Jan-May 2022

I reproduced and extended the proposed Genetic Algorithm; which focuses on replicating the conditions of grouping and dispersal in bacterial micro-colonies to provide insight into the evolution of cooperation (and further to understand if strategies exist to ensure cooperation is selected for in evolving societies). The extension focused on better representing individuals to allow for imprecisions in their understanding of the simulated environment. The paper and adjoint code can be found here.

Inspecting Functional Modularity in NNs

Reproduction of ICLR 2021 Paper - Jan-June 2022

A reproduction of, Are Neural Nets Modular? Inspecting Functional Modularity through Differentiable Weight Masks , R Csordas et Al., International Conference on Representation Learning, 2021. We reproduce the proposed tool for inspecting functional modularity and tested it on a range of NNs (notably CNNs and RNNs). We were ultimately able to replicate several results, though found the original implementation to be somewhat contrived.

The Importance of Group-Size Preferences for the Evolution of Cooperation Under the Conditions of Individual Selection

Literature Review and Model Extension - Jan-May 2022

I reproduced and extended the proposed Genetic Algorithm; which focuses on replicating the conditions of grouping and dispersal in bacterial micro-colonies to provide insight into the evolution of cooperation (and further to understand if strategies exist to ensure cooperation is selected for in evolving societies). The extension focused on better representing individuals to allow for imprecisions in their understanding of the simulated environment. The paper and adjoint code can be found here.

Online Trend Detection at Scale

MEng Group Design Project - Client: Senseye Ltd. - Sept-Jan 2021/22

I retained the position of Secretary and Primary Author for this group project.

We researched and designed three innovative generalised (online) models which, given a (very) large set of univariate time-series data, were capable of detecting numerous types of trends. I personally implemented a trend-detection model which focused on re-defining the notion of a trend as 'variance-related errors with respect to the current state of data'. This allowed us to comprehensively evaluate both monotonic and heteroscedastic trends, which reduces the loss of information through model selection bias and is capable of detecting various types of behaviour (which both classical and SoA approaches do not). In addition, I contributed a (novel) optimisation technique; rather than running the same evaluation processes on every data point, we ran a change-point detection algorithm (initially detecting trend-incidents rather than entire windows) to signal 'interesting' windows of data which we then evaluated using the usual evaluation process. This reduced computation by >30%. Documentation can be found here.

Generalised Ethical Dilemma Solver

MEng Thesis - Sept-May 2020/21

For my thesis I implemented a philosophical framework by relating classical ML techniques to philosophical concepts. The basis of this model was to prove that ML structures can reason (and provide reason) so long as the methods implemented match the dynamics of chosen philosophies. In addition, I demonstrated this application with respect to the generalised scope of ethical dilemmas which granted me the opportunity to test my hypothesis over various sociological and psychological situations where our agent would have to make a crucial decision. I got several professionals working in the fields of Psychology, Ethics and AI to evaluate the model and ensure that it does in fact reason (no technical testing methods by me would work here without compromising the integrity of the agent). As a result this dissertation was given the title of "award worthy" by the University of Southampton and suggested for publication. Source and Documentation can be found here.

Spatial Distribution of Sensor Network under Star Topology

Research Internship at Lurtis Rules - July-Nov 2020

During this internship I wrote a comparative paper (unpublished) on evaluating linear and non-linear strategies of node distribution for generalised irregular (agricultural) fields; supervised by Jose M. Pena. We worked with strategies concerning Differential Evolution and found that linear startegies can sometimes outperform non-linear strategies as a result of over-complication particular when presented with a nearly-regular "iregular" field shape.

Adrian Azzarelli

Adrian Azzarelli

About

Projects