You can add nonbinary weights and nonbinary level of cell activation. Then different predictions will have different levels of activation. So, you will be able to feed not a union of predictions but only the most likely(more active). This can solve the problem