I am using mgcv to train GAM. For example, suppose we have a Dataset with 3 Features: x, y and z. When should I use a Thin Plate Spline for each variable? When should I use a Thin Plate Spline with two or more Features as input (for example s (x, y))? I was thinking in cases where we had Data with Type 2-D or higher, for example in RGB Images Or if there are features with high correlation, is this a fair reasoning?