PolyNet: A Pursuit of Structural Diversity in Very Deep Networks

By X. Zhang et al

Published on July 17, 2017

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. PolyInception and PolyNet
3.1. PolyInception Modules
3.2. An Ablation Study

Summary

PolyNet is a study on exploring structural diversity in designing deep networks. It introduces PolyInception modules as building blocks for constructing very deep networks, aiming to enhance expressive power while maintaining computational cost. Various PolyInception structures are discussed, such as poly-2, mpoly-2, and 2-way. An ablation study is conducted using a trimmed down version of Inception-ResNet-v2 as the baseline model (IR 3-6-3). Different variants of IR 3-6-3 are created by replacing Inception residual units with PolyInception modules to evaluate their impact on performance.