Feature selection aims at reducing the dimensionality of the data by removing redundant or irrelevant features. It can reduce training time, improve explainability, and also enhance model performance in many machine learning tasks. In the literature, numerous feature selection algorithms have been proposed with various strengths and applications. A recently developed method, called the concrete autoencoder, provides a powerful tool for feature selection using a novel combination of autoencoders and a relaxation technique based on simulated annealing. A drawback of this method is that the number of features to be selected has to be determined in advance and can only be tuned by repeating the full feature selection process. The task of the student is to explore the concrete autoencoder method and propose a modification that can be implemented so that the autoencoder can adaptively change the number of selected features.