Interpretable Embedding Procedure Knowledge Transfer via Stacked Principal Component Analysis and Graph Neural Network

Knowledge distillation (KD) is one of the most useful techniques for\nlight-weight neural networks. Although neural networks have a clear purpose of\nembedding datasets into the low-dimensional space, the existing knowledge was\nquite far from this purpose and provided only limited information. We argue\nthat good knowledge should be able to interpret the embedding procedure. This\npaper proposes a method of generating interpretable embedding procedure (IEP)\nknowledge based on principal component analysis, and distilling it based on a\nmessage passing neural network. Experimental results show that the student\nnetwork trained by the proposed KD method improves 2.28% in the CIFAR100\ndataset, which is higher performance than the state-of-the-art (SOTA) method.\nWe also demonstrate that the embedding procedure knowledge is interpretable via\nvisualization of the proposed KD process. The implemented code is available at\nhttps://github.com/sseung0703/IEPKT.\n