This repository includes all code needed for the experiments illustrated in the MSc Dissertation 'Empirical Studies of Mixup and the Induced Training Strategy for PAC-Bayesian Learning' at UCL.
Synthetic Data Code file contains the code needed for experiments in the first part fo the project, which creates synthetic data to analyze the effect of data size, dimensionality, label corruption level, class overlapping level on the performance of mixup compared to ERM.
Real-World Data file contains the code needed for running mixup & ERM on MNIST (MLP), CIFAR-10 (ResNet18), CIFAR-100 (ResNet18) datasets.
PAC-Bayesian Learning file contains the code used in the second part of the project, using mixup as a training objective for prior in PAC-Bayesian learning that uses probabilistic neural network.