Portfolio

Portfolio

• Counting Temporal Repetitions
  • A fully self-supervised approach to repetition counting using no ground truth data, to estimate the labels of repetitive video datasets. (code currently unavailable)
• Contrasting Temporal Neighbours
  • A project using contrastive learning to improve action recognition performance on untrimmed videos. By contrasting temporal neighbours, we can improve model performance, where similar neighbours are often misclassified for one-another. (code currently unavailable)
• The Blind Camera
  • A research project investigating the utility of audio as a modality for egocentric action recognition. Incorporated multiple, (at the time) state of the art CNN approaches to audio classification, applying them to the novel scenario of egocentric action recognition. At the time of this project, there was no prior work performing egocentric audio-only action recognition
• P.I.E.S: Passive Information Extraction System
  • A collaborative project for the insurance company, LV. We developed a question answering system that transcribed live speech data and queried it with a fine-tuned language model to automate the collection of data, allowing the phone operator to focus on communicating with the person on the phone, instead of noting down their details. The project included speech-to-text, fine-tuning LLMs, online learning, allowing the model to improve over time, and a javascript front-end. This project has been adapted for internal use at LV.
• Fisher Distillation
  • My MSc dissertation project. An approach to scaling up Differentiable Architeure Search models for network distillation using Fisher information as a sensitivity metric. DARTS models are expanded using Fisher information to determine which cell to grow, and the highest scoring cells are expanded by reducing the number of grouped convolutions in that cell.
• Exploring CNNs
  • A CNN implementation in NumPy and PyTorch scripts for experimental work comparing striding and dilation. Coursework for Edinburgh University’s Machine Learning Practical module.
• Exploring RMSProp and Adam with weight decay
  • RMSProp, Adam, cosine annealing and weight decay implementations inspired by Fixing Weight Decay Regularization in Adam. Coursework for Edinburgh University’s Machine Learning Practical module.
• HMM Viterbi
  • An implementation of the Viterbi algorithm for HMM’, applied to POS tagging.
• Data Mining Over Cancer Data
  • A decision support tool to aid clinicians in the diagnosis of bowel cancer. Uses Weka to generate a selection of models and C# is used to parse the models and generate classifications and other metrics. My final year project at the University of Hull. Due to the medical nature of the data it cannot be made available. Consequently, the public version of this project is not functional.
• Twitter Sentiment Analysis
  • Live graphing of twitter sentiment analysis. Graph shows a live feed of the sentiment of tweets that include a specified keyword.
• Perceptron
  • A simple perceptron developed in Python/Numpy to predict a robot’s movements.
• Targeting Agent
  • An MLP to predict if a point is on target. Developed in Keras/Tensorflow
• Poker Hand Predictor
  • Predicting poker hands with Keras and Tensorflow. It uses the Poker Hand dataset from UCI. A group project from the University of Hull Deep Learning Winter School