BOIL

BOIL: Towards Representation Change for Few-shot Learning

Introduction

Name:	BOIL
Embed.:	Conv64F/ResNet12/ResNet18
Type:	Meta
Venue:	ICLR'21
Codes:	BOIL

‼️ The authors of BOIL didn't release the train args except for CIFAR-FS. We can't reproduce by the hyper-parameters they claimed. So we choose our hyper-parameters.

Cite this work with:

@inproceedings{DBLP:conf/iclr/OhYKY21,
  author    = {Jaehoon Oh and
               Hyungjun Yoo and
               ChangHwan Kim and
               Se{-}Young Yun},
  title     = {{BOIL:} Towards Representation Change for Few-shot Learning},
  booktitle = {9th International Conference on Learning Representations, {ICLR} 2021,
               Virtual Event, Austria, May 3-7, 2021},
  publisher = {OpenReview.net},
  year      = {2021},
  url       = {https://openreview.net/forum?id=umIdUL8rMH},
}

---

Results and Models

Classification

	Embedding	📖 miniImageNet (5,1)	💻 miniImageNet (5,1)	📖miniImageNet (5,5)	💻 miniImageNet (5,5)	📝 Comments
1	Conv64F	*49.61 ± 0.16	47.92 ± 0.35 ⬇️ 📋	*66.45 ± 0.37	64.39 ± 0.30 ⬇️ 📋	Table.2
2	ResNet12	-	58.87 ± 0.38 ⬇️ 📋	-	72.88 ± 0.29 ⬇️ 📋	Table.2
3	ResNet18	-	57.85 ± 0.40 ⬇️ 📋	-	70.84 ± 0.29 ⬇️ 📋	Table.2

	Embedding	📖 tieredImageNet (5,1)	💻 tieredImageNet (5,1)	📖tieredImageNet (5,5)	💻 tieredImageNet (5,5)	📝 Comments
1	Conv64F	*48.58 ± 0.27	50.04 ± 0.38 ⬇️ 📋	*69.37 ± 0.12	65.51 ± 0.34 ⬇️ 📋	Table.2
2	ResNet12	-	64.66 ± 0.44 ⬇️ 📋	-	80.38 ± 0.31 ⬇️ 📋	Table2
3	ResNet18	-	60.85 ± 0.45
⬇️ 📋	-	77.74 ± 0.38 ⬇️ 📋	Table.2