Improving Pre-Trained Self-Supervised Embeddings Through Effective Entropy Maximization