Relation Networks for few-shot learning

E899065

deep learning model few-shot learning method metric-based meta-learning method

Relation Networks for few-shot learning is a deep learning approach that learns a generic, trainable similarity function to compare query and support examples for few-shot classification tasks.

Jump to: Statements Referenced by

Statements (47)

Predicate	Object
instanceOf	deep learning model ⓘ few-shot learning method ⓘ metric-based meta-learning method ⓘ
addressesTask	few-shot classification ⓘ one-shot classification ⓘ
assumes	small labeled support set per novel class ⓘ
citationVenue	Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition NERFINISHED ⓘ
classificationDecisionBy	argmax over relation scores ⓘ
comparedTo	Matching Networks NERFINISHED ⓘ Prototypical Networks NERFINISHED ⓘ
compares	query examples ⓘ support examples ⓘ
doesNotRequire	fine-tuning on novel classes at test time ⓘ
embeddingModuleType	convolutional neural network ⓘ
evaluatedOn	Omniglot NERFINISHED ⓘ miniImageNet NERFINISHED ⓘ
evaluationSetting	N-way K-shot classification ⓘ
field	computer vision ⓘ machine learning ⓘ meta-learning ⓘ
hasCoreIdea	learn a deep, trainable similarity function between query and support examples ⓘ
hasFullName	Relation Network for Few-Shot Learning NERFINISHED ⓘ
improvesOn	hand-designed distance metrics ⓘ
inspiredFollowUpWork	relation-based few-shot detection methods ⓘ relation-based few-shot segmentation methods ⓘ
introducedInPaper	Learning to Compare: Relation Network for Few-Shot Learning NERFINISHED ⓘ
keyContribution	jointly learn embeddings and similarity function end-to-end ⓘ
learningParadigm	supervised meta-learning ⓘ
learningType	inductive learning from few examples ⓘ
notableProperty	simple architecture with strong few-shot performance ⓘ
optimizationMethod	stochastic gradient descent ⓘ
outputs	relation scores between query and each support class ⓘ
proposedBy	Flood Sung NERFINISHED ⓘ Li Zhang NERFINISHED ⓘ Philip H. S. Torr NERFINISHED ⓘ Tao Xiang NERFINISHED ⓘ Timothy M. Hospedales NERFINISHED ⓘ Yongxin Yang NERFINISHED ⓘ
publishedAtConference	CVPR 2018 NERFINISHED ⓘ
publishedInYear	2018 ⓘ
relationModuleType	neural network that outputs similarity scores GENERATED ⓘ
relationScoreRange	[0,1] similarity score GENERATED ⓘ
trainedWith	episodic training strategy ⓘ
trainingMimics	few-shot evaluation episodes ⓘ
usesComponent	embedding module ⓘ relation module ⓘ
usesLoss	mean squared error on relation scores ⓘ

Referenced by (1)

Full triples — surface form annotated when it differs from this entity's canonical label.

Matching Networks for One Shot Learning → influenced → Relation Networks for few-shot learning ⓘ