Beyond Black Box Densities: Parameter Learning for the Deviated Components

Most data-driven learning processes typically consist of iterative steps that involve model training and fine-tuning, with more data in-take leading to further model re-training and refinement. As more samples come in and exhibit more complex patterns, the initial model may be obsolete, risks being discarded or absorbed into a richer class of models to adapt better to increased complexity. It takes much resources to train complex models on a rich data population. Moreover, many successful models in modern real-world applications have become so complex that make them hard to properly evaluate and interpret; aside from the predictive performance they may as well be considered as black boxes. Nonetheless, as data populations evolve and so must the learning models, several desiderata remain worthy: the ability to adapt to new complexity while retaining aspects of old "wise" model, and the ability to interpret the changes.