PROJECT TITLE :

Unsupervised Rotation Factorization in Restricted Boltzmann Machines

ABSTRACT:

Computer vision relies heavily on the ability to select the right images for the task at hand. Rotation-invariant feature learning has recently been proposed using several extensions to the original Restricted Boltzmann Machine (RBM) model. Rotation nuisance in 2D picture inputs is factorised explicitly using an unique extended RBM that we offer in this paper to learn rotation invariant features. When training, our model uses information about the reconstruction error to estimate the orientation of each image, rather than trying to learn invariant features. A Kullback-Leibler divergence provides stability and uniformity to the training process. The -score, a measure of invariance, was utilised to demonstrate that our approach learns rotation-invariant features analytically and experimentally. Using the test accuracy of an SVM classifier, we demonstrate that our method beats the existing state-of-the-art RBM approaches for rotation invariant feature learning.