Implementation of a Gaussian Process regression from scratch and proposed a way to approximate the hard computation (of the update) of the model posterior. Then applied to an inference problem based on space data.
Coding exercise based on the theory shown in Variational Inference for Neural Networks, implementing a simple Bayesian NN, and applying it on the Rotated MNIST and Fashion MNIST datasets.
Implementation of a custom Bayesian optimization algorithm to an hyperparameter tuning problem. In particular, we wished to determine the value of a model hyperparameter that maximizes the validation accuracy subject to a constraint on the average prediction speed, as shown in Bayesian Optimization with Unknown Constraints for the general case.
The task was to implement an algorithm that, by practicing on a simulator, learns a control policy for a lunar lander. The method suggested is a variant of policy gradient with two additional features, namely (1) Rewards-to-go, and (2) Generalized Advantage Estimatation, both aiming at decreasing the variance of the policy gradient estimates while keeping them unbiased.