Lecture 11

Post date: May 08, 2017 1:33:29 PM

We made a review of the SVM and then followed two ideas

1. Random, untrained weights can work surprisingly well (and can sometimes be used e.g. for architecture search). Notebooks: Echo State Networks (recurrent network with untrained random weights): https://github.com/janchorowski/nn_assignments/blob/master/lecture_notebooks/09-ESNmackey.ipynb random weights in a two-layer network: https://github.com/janchorowski/nn_assignments/blob/master/lecture_notebooks/11_random_weights.ipynb

2. We have seen the Variational Autoencoder (paper: https://arxiv.org/abs/1312.6114, demo: https://github.com/janchorowski/nn_assignments/blob/master/lecture_notebooks/11_aevb.ipynb)