It can use both labeled data (data with explanations) and unlabeled data to improve the accuracy of its feature extraction.
Traditional methods often use the Hirschfeld-Gebelein-Rényi (HGR) maximal correlation, which is powerful but requires strict mathematical "whitening" constraints. These constraints make the math very difficult to calculate and unstable during training. 6585mp4
Soft-HGR relaxes these "hard" constraints into a "soft" objective. It uses a straightforward calculation involving just two inner products, making the process much faster and more stable. Key Features and Benefits It can use both labeled data (data with