Supervised learning is a broad approach in statistical learning. The model learns to map an input to an output based on example input-output pairs.
Some supervised approaches include regression (continuous data) and classification (discrete data). Many possible approaches may be suitable — for instance when fitting a regression model, a higher-degree model may fit the data better.
The sweet spot is to avoid underfitting (no learning) nor overfitting (memorisation). This allows the model to generalise well.
The choice of which approach to use is dependent on the problem. The inductive bias (or learning bias) is a set of human assumptions for a model’s architecture and algorithm to help us solve our problem. For instance, MLPs have an inductive bias towards composing an approximation of a complicated function with simpler functions.
Some challenges: requires a large amount of labelled data (which can be expensive to label and to obtain), and most data isn’t labelled anyways (Internet data).