meta-transfer learning for few-shot learningyaliu/files/meta-transfer... · 2020-07-14 ·...
TRANSCRIPT
Meta-Transfer Learning for Few-Shot LearningQianru Sun1,3* Yaoyao Liu1,2* Tat-Seng Chua1 Bernt Schiele3
1National University of Singapore, 2Tianjin University, 3Max Planck Institute for Informatics
Motivation & Contributions Meta-Transfer Learning Hard Task Meta-Batch● Few-shot learning is challenging
due to the lack of training data.● Re-thinking promising methods:○ deep neural networks (DNN)○ transfer learning (pre-train, fine-tune)○ meta-learning (meta gradient descent)
● Problem: “catastrophic forgetting”
● Our solution: Scaling & Shifting (SS)
● Problem: slow meta-training convergence
● Our solution: hard negative task sampling
Detail
Top performance is achieved!
Faster convergence is achieved!
Detail
● miniImageNet dataset
Few-shot CIFAR-100 (FC100) dataset
method backbone 1-shot 5-shot
● Few-shot CIFAR-100 (FC100) datasetmethod backbone 1-shot 5-shot
★ Code is available at: https://github.com/y2l/meta-transfer-learning-tensorflow