This project was created for the Laboratory of Computational Physiscs of the Physiscs of Data master program, and it can also be found in the threeblueonebrowneyes github organization, since is the shared work of all the authors. The complete code is available in the 'group2305_exercise2.ipynb' file. The report on the RBM is presented in the 'group2305_assignment.pdf' file, and the abstract is the following:
Nowadays, model explainability in Machine Learning algorithms is an issue often overlooked. However, in some circumstances, a human-understandable explanation of algorithm decision-making behavior is required. Typical examples concern the understanding of patterns hidden in the data, where a standard statistical analysis would fail. In order to showcase one model, we test Restricted Boltzmann Machines as a tool to discover already known correlations in simplified protein structures. We test the explainability of RBMs by checking the recurring patterns appearing in the trained weights. We also verify the RBM susceptibility to the tuning of its hyperparameters: the number of hidden units, the encoding representation, CD-n steps, training and generating amplitudes.