Learning Representations by Maximizing Mutual Information Across Views