Notable Projects

Neural network models for influenza forecasting with associated uncertainty using Web search activity

In this work I used Bayesian recurrent neural networks to forecast influenza-like illness (ILI) in the United States. The methodological framework improves over the state-of-the-art in 2 forecasting metrics. The model achieves this by using Web search activity time series in conjunction with historical ILI rates as observations for training neural network (NN) architectures.

Forecasts for Iterative Recurrent Neural Network (IRNN)

Iterative Recurrent Neural Network (IRNN) Diagram

Forecasting Influenza-like Illness using Physics Informed Neural Ordinary Differential Equations

paper

github

While working as a researcher at UCL I developed a (currently unpublished) method for using Physics informed neural ODEs to forecast influenza at a state, regional, and national level in the US. The model is based on a universal differential equations and uses: a compartmental model who's parameters are estimated by a neural network, a separate neural network to reduce error in the model outputs, and is wrapped in a VAE framework which estimates the initial conditions from ILI rates and web search data.

European Robotics League Challenge Seville

Winner of SESAR prize for multirotor collision avoidance. I co-lead a research team to develop a multi rotor for autonomous navigation of environments using depth sensing camera and GPS. It was later modified to use visual SLAM and operate indoors.

Vision System for Team Bath Drones

github

I built a first of its kind target recognition system for a fixed wing drone to identify, geolocate, and relay the position of alphanumeric targets. The system uses a OpenCV and a convolutional neural network to identify and find the position of targets in a frame, calculates the position of the target from the drones GPS position, attitude and the position of the target in the image. Targets are clustered to improve accuracy and give an estimate of uncertainty.

UAV pre flight.

UAV control and vision computer.

An Nvidia Jetson Nano (top) runs image recognition software, taking position and attitude data from the Pixhawk flight controller. (centre)

UAV Internals.

Flight controller and vision system computer (top), camera and canard mechanism (bottom)

Vision system test aircraft.

Multirotor for easy testing of the vision system.

UAV in flight.

Smart Cities Robotics Challenge Milton Keynes

Winner of the SciRoc Challenge for multirotor collision avoidance indoors. I co-lead a research team where we developed a collision avoidance algorithm using SLAM for positioning and depth sensing camera for path planning.