Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification

TLDR

Rectified activation units are essential for state‑of‑the‑art neural networks. The study aims to improve image classification by introducing a Parametric Rectified Linear Unit and a rectifier‑aware initialization method for rectifier neural networks. The authors propose PReLU and a robust initialization that enables training extremely deep rectified models from scratch. PReLU networks attain a 4.94 % top‑5 test error on ImageNet 2012, a 26 % relative improvement over the 2014 winner and the first to beat human‑level performance, with negligible extra computation and low overfitting risk.

Abstract

Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on our PReLU networks (PReLU-nets), we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%). To our knowledge, our result is the first to surpass human-level performance (5.1%, Russakovsky et al.) on this visual recognition challenge.