[2Win5] Poster session 2

[2Win5-56] Feature generalizability in the intermediate layer of CNNs and its application to transfer learning

〇Mizuki Dai¹, Kenya Zin'no¹ (1.Tokyo City University)

Keywords:Convolutional Neural Network, Class Classification, Transfer Learning

This study focuses on the changes in feature representations across the layers of Convolutional Neural Networks (CNNs), with particular emphasis on whether the shallow layers extract generic features that are independent of the training dataset. Centered Kernel Alignment (CKA) was employed as the evaluation metric to quantitatively assess the similarity of feature representations in each layer of CNN models trained on different datasets. The results suggest that, in the shallow layers, common features that do not depend on the training dataset are maintained. Furthermore, with the aim of optimizing transfer learning, it was confirmed that applying additional training (fine-tuning) to the layers with low CKA similarity enhances classification accuracy. The outcomes of this study contribute not only to a deeper understanding of the internal structure of CNNs but also to the establishment of efficient training strategies in transfer learning.