Multi-Task Learning with Knowledge Transfer for Facial Attribute Classification

Having achieved satisfying performance in multiple areas, multi-task learning (MTL) is being applied on facial attribute classification. However, most multi-task learning algorithms neglect the latent connections among facial attributes, sorting attributes based on local information only, or merely viewing each attribute as independent. The concept of curriculum learning suggests that in multi-task learning, "easy" tasks can be learned first and used to guide the learning process of "hard" tasks. Inspired, we propose KT-MTL, a novel MTL network with knowledge transfer for facial attribute classification. Depending only on label information, attributes are divided into multiple tasks by spectral clustering and labeled as "strong" or "weak" embodying their correlation extent. During training, parameters learned in "strong" network are transferred to "weak" net, imitating the teacher-student learning process. Both parts contribute to the total loss with a specifically designed loss function. The proposed network archives a competitive overall accuracy score of above 92% on aligned CelebA images and the highest accuracy of 91.89% on "weak" tasks.