Abstract
A key problem in deep multi-attribute learning is to effectively discover theinter-attribute correlation structures. Typically, the conventional deepmulti-attribute learning approaches follow the pipeline of manually designingthe network architectures based on task-specific expertise prior knowledge andcareful network tunings, leading to the inflexibility for various complicatedscenarios in practice. Motivated by addressing this problem, we propose anefficient greedy neural architecture search approach (GNAS) to automaticallydiscover the optimal tree-like deep architecture for multi-attribute learning.In a greedy manner, GNAS divides the optimization of global architecture intothe optimizations of individual connections step by step. By iterativelyupdating the local architectures, the global tree-like architecture getsconverged where the bottom layers are shared across relevant attributes and thebranches in top layers more encode attribute-specific features. Experiments onthree benchmark multi-attribute datasets show the effectiveness and compactnessof neural architectures derived by GNAS, and also demonstrate the efficiency ofGNAS in searching neural architectures.