Implementation of a Gaussian Process regression applied on a real world inference problem (ground-water pollution prediction). Nyström approximation of the kernel function and inducing points method were applied to cope with the computational burden of the problem, due to the very large dataset.
Implementation of a Bayesian Neural Network, trained and tested on the Rotated MNIST and Fashion MNIST datasets, for class prediction with uncertainty. The training process minimizes a loss function which considers both a Cross-Entropy loss and a Kullback–Leibler divergence loss term.
Implementation of a Bayesian Optimization algorithm that performs hyperparameter tuning with constraints. Expected Improvement (EI) and Upper Confidence Bound (UCB) activation functions were both deployed to implement the algorithm.
Implementation of a Deep Reinforcement Learning algorithm able to learn a control policy for a lander (spaceship), by practicing on a simulator. The project required the use of Actor-Critic methods with policy gradients, in particular Rewards-to-go and Generalized Advantage Estimatation methods, both aiming at decreasing the variance of the policy gradient estimates.
Video example: