Learning Local Receptive Fields in Deep Belief Networks for Visual Feature Detection

Through the introduction of local receptive fields, we improve the fidelity of restricted Boltzmann machine (RBM) based representations to encodings extracted by visual processing neurons. Our biologically inspired Gaussian receptive field constraints encourage learning of localized features and can seamlessly integrate into RBMs. Moreover, we propose a method for concurrently finding advantageous receptive field centers, while training the RBM. The strength of our method to reconstruct characteristic details of facial features is demonstrated on a challenging face dataset.