This repository contains typeset reports for some of my favourite projects I have done during my studies.

I describe the projects I have included below.

Some of the projects are a result of teamwork and I have mentioned my collaborators in the descriptions of those.

As a part of the Machine Learning Practical course in the University of Edinburgh, Miroslav Trifonov, Kamen Brestnichki and I designed and implemented a differentiable model that can solve a class of simple visual IQ questions called Raven Progressive Matrices.

As part of the course, we had to produce two reports. The first one motivates our choice of task and dataset and contains some preliminary experiments and the second one describes our method and experiments. Our code can be found here

Under the supervision of Dr Iain Murray

Dasher is a continuous-gesture text-entry system developed in Cambridge in 2005. It is particularly suitable for people with certain disabilities and is used by a small minority of people. For its operation it relies on a character-level language model -- a model that outputs a probability for the next character in a sequence given all of the preceding characters.

At the time of writing, the state-of-the-art in language modelling was achieved using Recurrent Neural Network-based language models. For my project, I investigated the sensibility and feasibility of the idea of incorporating this type of language model in the Dasher system.

I designed and carried out experiments that compared RNN-based models to the language model currently employed in Dasher in terms of language modelling capacity as well as computational requirements.

My thesis can be found here and my code and experimental results here and here.

Under the supervision of Dr Iain Murray

My Master thesis built on the work I did for my Bachelor thesis described above.

A feature currently available in Dasher is that the language model can adapt to a specific user's writing style, meaning they will be able to enter text more quickly as this happens.

Adapting RNN-based models is an open research area and there are challenges connected to implementing adaptive models in Dasher. In my work I compare existing approaches to adapting neural models and also propose an approach that improves the initial predictions in Dasher.

The thesis can be found here and my code and experimental results here and here.

University of Edinburgh, Academic year 2018/2019, taught by Prof. Sharon Goldwater and Henry Thompson

In collaboration with Marko Smilevski

For this assignment we were tasked with exploring problems with distributed word embeddings asn similarity metrics. We chose to look into the hubness problem -- the phenomenon that often times when measuring similarities between words using distributed embeddings we see that some high-frequency words are "closer" to a disproportionally large number of words.

We explore several different methods for producing embeddings and computing similarities. We end up with some interesting findings which possibly explain some claims we have found in literature on the subject. Our rather short (due to a hard limit in the assignment formulation) report can be found here

I had to build and parallelize an agent for a slightly modified version of the classical board game Othello (also known as Reversi).

A rather long report, detailing my work can be found here