Preventing Model Collapse in Deep Canonical Correlation Analysis by Noise Regularization