Generating images with recurrent adversarial networks
Daniel Jiwoong Im1 Montreal Institute for Learning Algorithms University of Montreal Chris Dongjoo Kim Department of Engineering and Computer Science York University Hui Jiang Department of Engineering and Computer Science York University Roland Memisevic Montreal Institute for Learning Algorithms University of Montreal
Abstract
Gatys et al. (2015) showed that optimizing pixels to match features in a convolutional network is a way to render images of high visual quality. Unrolling this gradient-based optimization can be thought of as a recurrent computation, that creates images by incrementally adding onto a visual “canvas”. Inspired by this view we propose a recurrent generative model that can be trained using adversarial training. In order to quantitatively compare adversarial networks we also propose a new performance measure, that is based on letting the generator and discriminator of two models compete against each other.
中文速览
对抗生成网络（GAN）在生成逼真图像方面表现出色，但单步生成方式难以捕捉复杂的数据分布，且缺乏客观的定量评估手段。受卷积网络特征匹配式图像渲染与 DRAW 循环生成模型的启发，作者提出了生成式循环对抗网络（GRAN），让生成器以循环方式在多个时间步上把噪声逐步"画"到同一张画布上，最终叠加出完整图像，整个过程由对抗训练驱动而无需注意力机制或变分目标。为了解决 GAN 难以定量比较的痛点，作者还设计了一种"交叉对战"评估方案，让两个模型的生成器和判别器互相交叉对抗，以胜负关系衡量相对优劣。实验表明 GRAN 在视觉质量上优于基线 GAN，而交叉对战指标也为今后比较各类对抗生成模型提供了一种更客观、可复现的标准。
原文 arXiv:1602.05110;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1602.05110v5