Learning Optimal Representations with the Decodable Information Bottleneck