Graph-based Isometry Invariant Representation Learning : Supplementary material

We use supervised learning and train our network so that it maximizes the log-probability of estimating the correct class of training samples via logistic regression. Overall, we need to compute the values of the parameters in each convolutional and in fully-connected layers. The other layers do not have any parameter to be estimated. We train the network using a classical back-propagation algorithm and learn the parameters using ADAM stochastic optimization (Kingma & Ba, 2014).