Knowledge Transfer Graph for Deep Collaborative Learning

Soma Minami, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020


Knowledge transfer among multiple networks using their outputs or intermediate activations have evolved through manual design from a simple teacher-student approach to a bidirectional cohort one. The major components of such knowledge transfer framework involve the network size, the number of networks, the transfer direction, and the design of the loss function. However, because these factors are enormous when combined and become intricately entangled, the methods of conventional knowledge transfer have explored only limited combinations. In this paper, we propose a novel graph representation called knowledge transfer graph that provides a unified view of the knowledge transfer and has the potential to represent diverse knowledge transfer patterns. We also propose four gate functions that control the gradient and can deliver diverse combinations of knowledge transfer. Searching the graph structure enables us to discover more effective knowledge transfer methods than a manually designed one. Experimental results show that the proposed method achieved performance improvements.

Related Material

[pdf] [arXiv] [code]
@InProceedings{Minami_2020_ACCV, author = {Minami, Soma and Hirakawa, Tsubasa and Yamashita, Takayoshi and Fujiyoshi, Hironobu}, title = {Knowledge Transfer Graph for Deep Collaborative Learning}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {November}, year = {2020} }