The increasingly popular application of AI runs the risk of amplifying social bias, such as classifying non-white faces as animals. Recent research has largely attributed this bias to the training data implemented. However, the underlying mechanism is poorly understood; therefore, strategies to rectify the bias are unresolved. Here, we examined a typical deep convolutional neural network (DCNN), VGG-Face, which was trained with a face dataset consisting of more white faces than black and Asian faces. The transfer learning result showed significantly better performance in identifying white faces, similar to the well-known social bias in humans, the other-race effect (ORE). To test whether the effect resulted from the imbalance of face images, we retrained the VGG-Face with a dataset containing more Asian faces, and found a reverse ORE that the newly-trained VGG-Face preferred Asian faces over white faces in identification accuracy. Additionally, when the number of Asian faces and white faces were matched in the dataset, the DCNN did not show any bias. To further examine how imbalanced image input led to the ORE, we performed a representational similarity analysis on VGG-Face's activation. We found that when the dataset contained more white faces, the representation of white faces was more distinct, indexed by smaller in-group similarity and larger representational Euclidean distance. That is, white faces were scattered more sparsely in the representational face space of the VGG-Face than the other faces. Importantly, the distinctiveness of faces was positively correlated with identification accuracy, which explained the ORE observed in the VGG-Face. In summary, our study revealed the mechanism underlying the ORE in DCNNs, which provides a novel approach to studying AI ethics. In addition, the face multidimensional representation theory discovered in humans was also applicable to DCNNs, advocating for future studies to apply more cognitive theories to understand DCNNs' behavior.

人工智能的日益流行的应用存在放大社会偏见的风险，例如将非白人面孔归类为动物。最近的研究很大程度上将这种偏差归因于所实施的训练数据。然而，人们对其背后的机制知之甚少；因此，纠正偏见的策略尚未解决。在这里，我们检查了一个典型的深度卷积神经网络（DCNN）VGG-Face，它是使用由白人面孔组成的面部数据集进行训练的，而黑人和亚洲面孔则多于白人面孔。迁移学习结果显示，在识别白人面孔方面表现明显更好，类似于人类众所周知的社会偏见，即其他种族效应（ORE）。为了测试这种效果是否是由于人脸图像不平衡造成的，我们用包含更多亚洲人脸的数据集重新训练了 VGG-Face，发现了一个反向的 ORE，即新训练的 VGG-Face 在识别准确率上更喜欢亚洲人脸而不是白人人脸。此外，当数据集中亚洲面孔和白人面孔的数量匹配时，DCNN 没有表现出任何偏差。为了进一步检查不平衡的图像输入如何导致 ORE，我们对 VGG-Face 的激活进行了表征相似性分析。我们发现，当数据集包含更多的白人面孔时，白人面孔的表示更加清晰，通过较小的组内相似性和较大的表示欧几里得距离来索引。也就是说，白色面孔在 VGG-Face 的表征面孔空间中比其他面孔分散得更稀疏。重要的是，人脸的独特性与识别准确度呈正相关，这解释了在 VGG-Face 中观察到的 ORE。 总之，我们的研究揭示了 DCNN 中 ORE 的机制，为研究人工智能伦理提供了一种新方法。此外，在人类中发现的面部多维表示理论也适用于 DCNN，提倡未来的研究应用更多的认知理论来理解 DCNN 的行为。