Learning Rotation-Aware Features: From Invariant Priors to Equivariant Descriptors Supplemental Material

The R-FoE model of Sec. 3 of the main paper was trained on a database of 5000 natural images (50 × 50 pixels) using persistent contrastive divergence [12] (also known as stochastic maximum likelihood). Learning was done with stochastic gradient descent using mini-batches of 100 images (and model samples) for a total of 10000 (exponentially smoothed) gradient steps with an annealed learning rate. We trained the model using conditional sampling to avoid boundary issues [8]. Both learned filters were initialized randomly from a standard normal distribution, and constrained to have mean 0 and norm 1 throughout learning. We initialized the shapes of the potential functions to be very broad (cf . [5]).