Projects

This project explored using various learning algorithms to predict trial outcomes using the transcripts of the Old Bailey proceedings. The dataset comprised of 25,000 trials labeled as guilty or not guilty. Five classifiers were implemented from scratch, including ID3, Perceptron, SVM with SSGD, logistic regression with SGD, and bagged ID3, and a neural network was implemented using TensorFlow/Keras. Different feature representations like bag-of-words, tfidf, glove embeddings, and categorical attributes were examined. Neural networks achieved the best performance (0.83 on Kaggle), highlighting their suitability for nonlinear data. Other algorithms performed comparably (0.79-0.80 on Kaggle). The project emphasized the importance of algorithm selection, feature engineering, representation, and hyperparameter tuning in machine learning.

Analyzed two methods for quantifying the predictive uncertiainty of DeformerNet, a framework that provides a learning-based approach for minipulating deformable objects to a desired goal shape