Learning Representations Using Causal Invariance