Learning What and Where to Transfer

• DOI: 10.48550/arxiv.1905.05901
• Publication Date: 2019-01-01
• Authors: Yunhun Jang; Hankook Lee; Sung Ju Hwang; Jinwoo Shin
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; 03 medical and health sciences; 0302 clinical medicine; Statistics - Machine Learning; Machine Learning (stat.ML); Machine Learning (cs.LG)

As the application of deep learning has expanded to real-world problems with insufficient volume training data, transfer recently gained much attention as means improving performance in such small-data regime. However, when existing methods are applied between heterogeneous architectures and tasks, it becomes more important manage their detailed configurations often requires exhaustive tuning on them for desired performance. To address issue, we propose a novel approach based meta-learning that can automatically learn what knowledge from source network where target network. Given networks, an efficient scheme meta-networks decide (a) which pairs layers networks should be matched (b) features how each feature transferred. We validate our meta-transfer against recent various datasets architectures, automated significantly outperforms prior baselines find "what transfer" hand-crafted manner.