I tried to implement the Restricted Boltzmann Machine to confirm the utility of the deep learning method. I implemented a RBM and fed the MNIST character recognition data for one layer reconstruction.
But the learned weights of edges were same for each hidden nodes.
then I got wondered why the RBM can obtain asymmetric filters from the symmetric network topology. I couldn't grasp the situation from mathematical formula on thesis around RBMs.
Give me an intuitive insight for this.
