But how is that different from other forms of adjusting your models: adding some latent variables, add models for the error or enriching a statistical model in other ways, or playing with network architectures and weights sharing in deep learning models, etc. Or even further: playing with regularization, feature engineering, data normalization, data…