I designed a Gaussian (Gaussian distributed visible layer) - Bernoulli (binary distributed) RBM model (for reference, see: Improved Learning of Gaussian-Bernoulli Restricted Boltzmann Machines, pdf) and trained it using MNIST handwritten digits. Now would like to understand the learned model through its weight values. In general, weight values of this model should be positive or negative. My model's weight values are in the range of -1.46 to 1.75. Is this reasonable?
Many research papers suggest (for reference, see: Visually debugging restricted Boltzmann machine training with a 3D example, pdf, Tracking process from this website) that when we visualize weight values of RBM model, it will look like strokes. But when I visualize them, I am not able to get stokes kind of weights. What might be a reason for this?
I would like to know what kind of weight (filter) values a model should have. I have attached the visualization from a MATLAB implementation and expected output from internet.