Transformers have millions of parameters, whereas when I use linear regression in scikit learn it learns one parameter, where would I use machine learning over transformers, if the number of parameters in machine learning is so less compared to transformer?