 Initialization Problem

Thanks for your greate jod. In my experiments, I find it is difficult to converge with prototypes randomly initialized. How can I solve this? I have no prior on the dataset. With a Softmax CNN, the convergence process is stable.