On the Generalization and Causal Explanation in Self-Supervised Learning