UCL COMP0171 Bayesian Deep Learning (2023/24)

This repository contains the first coursework I completed for my MSc module COMP0171 Bayesian Deep Learning.

Tasks for coursework 1

1. Bayesian Inference

   • Computed log joint probability of a bayesian model
   • Implemented a MCMC sampler for estimating the posterior
   • Short questions on the MCMC method

2. Bayesian Classifier

   • Implemented a Bayesian logistic regression model
     • Finding maximum a posterior (MAP) esitmate through gradient descent
     • Approximating the posterior using Laplace approximation
     • Comparing features sets by estimating the model evidence
   • Short questions on the feature set selection strategy and overfitting problem in a bayesian setting

Tasks for coursework 2

1. Bayesian Neural Network

   • Implemented stochastic gradient Langevin dynamics (SGLD) for sampling network parameters
   • Estimated epistemic and aleatoric uncertainty
   • Analysed temperature scaling and its impact on Expected Calibration Error (ECE)

2. Fitting a Variational Auto-Encoder (VAE)

   • Designed and implemented the encoder and decoder for the MNIST data
   • Trained the VAE to maximise per-datapoint evidence lower bound (ELBO)

Python Environment

Requirement: python=3.11

pip install -r requirements.txt