Suppose of MNIST data, if I only label once for every possible digit (10 digits) and leave the rest unlabeled. Then I train them with multi-task learning, where the first task is classification (only labeled data) and regression (both of which have labels and images).
What did that problem call? Unsupervised? Supervised? Self-supervised? Semi-supervised? Weak-supervised?
The training pattern which you have followed is called semi-supervised learning in terms of machine learning
Semi-supervised learning involves a mix of labeled and unlabeled data. In your case, you have a small amount of labeled data (one labeled example for each digit) and a large amount of unlabeled data (the rest of the MNIST dataset), making it a semi-supervised approach
The classification task is supervised for the labeled data, while the regression task involves both labeled and unlabeled data, potentially learning from the structure of the unlabeled examples. This fits well into the semi-supervised learning paradigm, where the model leverages the information from unlabeled data to improve its learning process.
