The aim of this assignment is to master modern tools (TensorFlow and Keras) for training deep neural networks. This paper is structured as follows: first, we provide a proof-of-understanding on three types of deep networks: convolutional neural networks, recurrent neural networks and autoencoders. For this, we studied and slightly adapted pre-existing tutorials and examples. For each of the three types, we describe the network, the problem we selected, the data, the experiment and the results. We also mention which adaptations were made to the original example.

Next, we apply deep learning to an original data set. In particular, a character-based RNN is trained on tweets from U.S. president Trump, and this network is be used to generate president Trump's 'next tweet'. Finally, this paper concludes with a brief discussion on this experiment, our findings and suggestions for further research.

Authors: Marije Sluiskes (Leiden University), Maxime Casara (Leiden University).