[论文解读] Women also Snowboard: Overcoming Bias in Captioning Models

Kaylee Burns, Lisa Anne Hendricks+3 more|arXiv (Cornell University)|Mar 26, 2018

Multimodal Machine Learning Applications参考文献 48被引用 120

一句话总结

本论文提出 Equalizer 模型及两种损失—Appearance Confusion Loss 与 Confident Loss—to reduce gender bias in image captioning by ensuring gender predictions rely on visual evidence and by adapting to varying gender distributions at test time.

ABSTRACT

Most machine learning methods are known to capture and exploit biases of the training data. While some biases are beneficial for learning, others are harmful. Specifically, image captioning models tend to exaggerate biases present in training data (e.g., if a word is present in 60% of training sentences, it might be predicted in 70% of sentences at test time). This can lead to incorrect captions in domains where unbiased captions are desired, or required, due to over-reliance on the learned prior and image context. In this work we investigate generation of gender-specific caption words (e.g. man, woman) based on the person's appearance or the image context. We introduce a new Equalizer model that ensures equal gender probability when gender evidence is occluded in a scene and confident predictions when gender evidence is present. The resulting model is forced to look at a person rather than use contextual cues to make a gender-specific predictions. The losses that comprise our model, the Appearance Confusion Loss and the Confident Loss, are general, and can be added to any description model in order to mitigate impacts of unwanted bias in a description dataset. Our proposed model has lower error than prior work when describing images with people and mentioning their gender and more closely matches the ground truth ratio of sentences including women to sentences including men. We also show that unlike other approaches, our model is indeed more often looking at people when predicting their gender.

研究动机与目标

• 识别字幕模型在训练数据中放大性别偏见的方式。
• 提出一个偏见缓解的字幕描述框架，促进基于合理原因的描述。
• 使性别预测依赖于视觉证据而非情境线索。
• 在训练集与测试集之间的分布转移下评估偏见的降低。
• 证明模型在预测性别词时关注的是人而非背景信息。

提出的方法

• 基础字幕框架使用 Image Features from InceptionV3 来初始化一个 LSTM 描述生成器。
• 两种新颖的损失：Appearance Confusion Loss (ACL) 与 Confident Loss (Conf) 使在存在性别证据时的字幕偏向性别证据，在证据缺失时减少对非证据线索的依赖。
• Appearance Confusion Loss 将在有证据的图像中移除性别信息，并在证据缺失时鼓励男性/女性词汇的等概率。
• Confident Loss 在性别证据存在时提高正确性别预测的置信度，使用基于商的置信度度量以允许性别中性词汇。
• 最终目标为 L = α L_CE + β L_AC + μ L_Con（实验中 α=1，β=10，μ=1）。
• 训练依赖于带有 ACL 的 ground-truth 性别推理遮罩的 MSCOCO-Bias 与 MSCOCO-Balanced 数据集。

实验结果

研究问题

• RQ1 字幕模型在预测带有性别词时是否会显现并放大性别偏见？
• RQ2 相较于基线，所提出的 ACL 与 Confident Loss 是否降低了性别词的错误分类率？
• RQ3 Equalizer 在测试时的分布转移下是否将字幕中的性别词分布与真实分布对齐？
• RQ4 Grad-CAM/显著性等解释是否显示模型在预测性别词时更关注人而非情境线索？
• RQ5 通过基于人物证据来对性别预测进行“对原因的正确性”判断，模型是否更符合“对的原因”原则？

主要发现

• Equalizer 在 MSCOCO-Bias 与 MSCOCO-Balanced 测试集上相较基线拥有最低的性别词错误率。
• 在 MSCOCO-Bias 上，Equalizer 的错误率为 7.02，低于所有消融模型和基线；在 MSCOCO-Balanced 上，Equalizer 的错误率为 8.10，也低于大多数变体。
• Equalizer 在各数据集上得到的性别比接近真实分布（Ratio Δ 值：在 MSCOCO-Bias 为 −0.03，在 MSCOCO-Balanced 为 0.13，适用于完整模型）。
• 消融结果显示 ACL 与 Confident Loss 是互补的；移除任一项都会降低性能（去掉 ACL 的 Equalizer 或者去掉 Conf 的模型错误更高）。
• Equalizer 使性别之间的结果差异（Jensen-Shannon 散度 0.018，在比较的模型中最低）。
• 视觉解释显示 Equalizer 在预测性别词时更常关注到人而非情境线索，支持“对的原因”原则。
• 在标注者置信度阈值下，Equalizer 在性别不明确时倾向使用性别中性词汇，在证据明确时使用性别化词汇，符合人类的类似模式。