Learning to Separate Domains in Generalized Zero-Shot and Open Set Learning: a probabilistic perspective