Composite Neural Network: Theory and Application to PM2.5 Prediction

This work investigates the framework and statistical performance guarantee of composite neural network, which is composed a collection pre-trained non-instantiated network models connected as rooted directed acyclic graph, for solving complicated applications. A model generally well trained, targeted to approximate specific function. The advantages adopting component in composing are two-fold. One benefiting from intelligence diligence domain experts, other saving effort data acquisition computing resources time training. Despite general belief that may perform better than any single component, overall characteristics not clear. In this work, we propose prove performs its components with high probability. study, explore application---PM2.5 prediction---to support correctness proposed theory. empirical evaluations PM2.5 prediction, constructed machine learning models.