Assignment solutions by @zkeal and @davidpomerenke for the Advanced Concepts of Machine Learning lecture held by Kurt Driessens at Maastricht University in fall 2020.
For solutions by other students, see @msvincognito/awesome-dke.
We implement the backpropagation algorithm for artificial neural networks, and apply it to a simple synthetic classification problem. Our implementation achieves an optimal encoding with a modest training time of only 50 epochs.
We use the Keras framework to create a more complex convolutional autoencoder model, and successfully apply it to image data. We explore the impact of different architectures. Finally, we try to apply our model to an image colorization problem.
We implement a Q-learning algorithm to solve the mountain car reinforcement learning problem. We observe good and robust results with a relatively simple implementation.
We review a prominent NeurIPS paper by researchers from DeepMind and the University of Berkeley, where they introduce a new reinforcement learning algorithm, Stochastic Latent Actor-Critic (SLAC). Their main contribution is to split the learning process into representation learning and task learning, thereby strongly improving performance.