做年龄无关的人脸识别 age-invariant face recognition (AIFR),需要将人脸表征分解为 identity-dependent & age-dependent components,解纠缠学习这两部分特征,且用互信息来监督学习到的特征;然后只使用身份特征进行人脸识别。 差分建模
- We obtain the age-related features through a FC layer from the initial features, and the identity-related features are obtained from the subtract between initial features and age-related features. 用减法得到解纠缠的特征表示
- 互信息估计时,使用Contrastive Log-ratio Upper Bound (CLUB),估计互信息的上界来进行最小化
- 可以画特征解纠缠前后的分布图来证明解纠缠学习效果

