Mortimer Sotom

Study of how to extend the application of Convolutional Neural Networks to graphs to design Graph Convolutional Networks. Tox21 dataset challenge. Molecular features of compounds were encoded at node level while the model performs graph level classification.
The dissertation provides an analysis of the literature to explain the underlying theory behind graph convolution.

To overcome the issue of the 'black-box' effect of Neural Networks and for the user to establish some level of… Show more

Study of how to extend the application of Convolutional Neural Networks to graphs to design Graph Convolutional Networks. Tox21 dataset challenge. Molecular features of compounds were encoded at node level while the model performs graph level classification.
The dissertation provides an analysis of the literature to explain the underlying theory behind graph convolution.

To overcome the issue of the 'black-box' effect of Neural Networks and for the user to establish some level of trust in the model, the LIME software helps provide an explanation as to why the model classified an instance/molecule as toxic or not. It identifies which molecular fragment of the molecule is responsible for its toxicity, information which can be interpreted by a field expert.

Research ongoing in spare time. Show less

Blockchain based voting platform with IBM Hyperledger for a democratic voting and transparent ballot counting system. The project also considered the necessary infrastructure and protocols required for a real life deployment scenario.

Features:
- Access permissions for different roles.
- Prevention of double-voting.
- Auditability and immutability of data.

Predict the trip duration of a taxi trip given a starting location and a destination. Machine Learning regression task from a Kaggle competition providing data of the NYC yellow taxis for 1 year.

Agent developed for a multi-lateral, multi-issue, time-based negotiation with uncertainty about the opponents' preferences with the aim of maximising its own utility while minimising the distance to the Nash equilibrium.

Photogrammetry is a technique able to generate 3D models of real objects from a particular set of pictures. This project explored the viability of using photogrammetry to display real objects into a virtual environments for projects such as creating virtual museums or virtual shops.

This was designed in Unreal Engine 4 and coded with the Blueprint Visual Scripting system. Model meshes were edited with ReMake and Blender.

Various softwares were compared from free mobile… Show more

Photogrammetry is a technique able to generate 3D models of real objects from a particular set of pictures. This project explored the viability of using photogrammetry to display real objects into a virtual environments for projects such as creating virtual museums or virtual shops.

This was designed in Unreal Engine 4 and coded with the Blueprint Visual Scripting system. Model meshes were edited with ReMake and Blender.

Various softwares were compared from free mobile applications to a state of the art software. A user guide was produced after numerous tests, providing object requirements, camera requirements and lighting requirements. Show less