Towards Shape Biased Unsupervised Representation Learning for Domain Generalization
Nader Asadi1 Amir M. Sarfi1 Mehrdad Hosseinzadeh2 Zahra Karimpour1 Mahdi Eftekhari1 1 Department of Computer Engineering, Shahid Bahonar University of Kerman, Iran 2 University of Manitoba {naderasadi, {a.m.sarfi,
Abstract
Shape bias plays an important role in self-supervised learning paradigm. The ultimate goal in self-supervised learning is to capture a representation that is based as much as possible on the semantic of objects (i.e. shape bias) and not on individual objects’ peripheral features. This is inline with how human learns in general; our brain unconsciously focuses on the general shape of objects rather than superficial statistics of context. On the other hand, unsupervised representation learning allows discovering label-invariant features which helps generalization of the model. Inspired by these observations, we propose a learning framework to improve the learning performance of self-supervised methods by further hitching their learning process to shape bias. Using distinct modules, our method learns semantic and shape biased representations by integrating domain diversification and jigsaw puzzles. The first module enables the model to create a dynamic environment across arbitrary domains and provides a domain exploration vs. exploitation trade-off, while the second module allows it to explore this environment autonomously. The proposed framework is universally adaptable since it does
中文速览
自监督学习（self-supervised learning）虽然能学到与标签无关的通用特征，但研究发现它和卷积神经网络一样，容易过度依赖纹理、局部统计等表面特征，而非物体的整体形状，导致模型在跨域场景下泛化能力有限。为此，本文提出一个结合"域多样化"（domain diversification）与"拼图预测"（jigsaw puzzle）的自监督学习框架：先用自适应实例归一化对每块拼图碎片随机替换风格/纹理，使纹理线索失效，迫使模型只能依靠形状信息来还原拼图顺序；同时引入探索-利用权衡机制，让模型在训练中动态游历任意域，从而学到域不变、形状偏置的表征。在PACS、Office-Home、VLCS、Digits等多个域泛化基准上的大量实验表明，该方法在大多数数据集上超越了现有最优方法。这项工作的意义在于指出并修正了自监督方法对表面统计量的内在偏见，为提升深度模型在真实世界域偏移下的鲁棒性提供了一条无需目标域标注的轻量化思路。
原文 arXiv:1909.08245;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1909.08245v2