Abstract
We introduce a unified, flexible, and easy-to-implement framework ofsufficient dimension reduction that can accommodate both linear and nonlineardimension reduction, and both the conditional distribution and the conditionalmean as the targets of estimation. This unified framework is achieved by aspecially structured neural network -- the Belted and Ensembled Neural Network(BENN) -- that consists of a narrow latent layer, which we call the belt, and afamily of transformations of the response, which we call the ensemble. Bystrategically placing the belt at different layers of the neural network, wecan achieve linear or nonlinear sufficient dimension reduction, and by choosingthe appropriate transformation families, we can achieve dimension reduction forthe conditional distribution or the conditional mean. Moreover, thanks to theadvantage of the neural network, the method is very fast to compute, overcominga computation bottleneck of the traditional sufficient dimension reductionestimators, which involves the inversion of a matrix of dimension either p orn. We develop the algorithm and convergence rate of our method, compare it withexisting sufficient dimension reduction methods, and apply it to two dataexamples.