Leveraging privileged information (PI), or features available during training but not at test time, has recently been shown to be an effective method for addressing label noise. However, the reasons for its effectiveness are not well understood. In this study, we investigate the role played by different properties of the PI in explaining away label noise. Through experiments on multiple datasets with real PI (CIFAR-N/H) and a new large-scale benchmark ImageNet-PI, we find that PI is most helpful when it allows networks to easily distinguish clean from noisy data, while enabling a learning shortcut to memorize the noisy examples. Interestingly, when PI becomes too predictive of the target label, PI methods often perform worse than their no-PI baselines. Based on these findings, we propose several enhancements to the state-of-the-art PI methods and demonstrate the potential of PI as a means of tackling label noise. Finally, we show how we can easily combine the resulting PI approaches with existing no-PI techniques designed to deal with label noise.

翻译：利用特许信息(PI),或者在培训期间提供但并非在测试时提供的特征,最近被证明是处理标签噪音的有效方法。然而,其有效性的原因并没有得到很好的理解。在本研究中,我们调查了PI的不同特性在解释标签噪音方面所起的作用。通过对使用真实的 PI(CIFAR-N/H)和新的大型基准图像网络-PI的多个数据集的实验,我们发现,如果PI能够让网络很容易地区分干净的数据和吵闹的数据,同时能够学习如何将吵闹的例子混为一谈的捷径。有趣的是,当PI对目标标签过于预测时,PI方法往往比其非PI基线要差。根据这些调查结果,我们建议对最新的PI方法进行若干次改进,并展示PI作为处理标签噪音的一种手段的潜力。最后,我们展示了我们如何轻松地将由此产生的PI方法与旨在处理标签噪音的现有非PI技术结合起来。</s>