SANE: Specialization-Aware Neural Network Ensemble

Ziyue Li, Kan Ren, XINYANG JIANG, Mingzhe Han, Haipeng Zhang, Dongsheng Li

Abstract: Real-world data is often generated by some complex distribution, which can be approximated by a composition of multiple simpler distributions. Thus, it is intuitive to divide the complex model learning into training several simpler models, each of which specializes in one simple distribution. Ensemble learning is one way to realize specialization, and has been widely used in practical machine learning scenarios. Many ensemble methods propose to increase diversity of base models, which could potentially result in model specialization. However, our studies show that without explicitly enforcing specification, pursuing diversity may not be enough to achieve satisfactory ensemble performance. In this paper, we propose SANE --- an end-to-end ensemble learning method that actively enforces model specification, where base models are trained to specialize in sub-regions of a latent space representing the simple distribution composition, and aggregated based on their specialties. Experiments in several prediction tasks on both image datasets and tabular datasets demonstrate the superior performance of our proposed method over state-of-the-art ensemble methods.

One-sentence Summary: We propose an end-to-end ensemble learning method to simultaneously achieve specialized base model learning and adaptive ensemble of base models.