Open letme-hj opened 1 year ago
ProtoNet (Prototypical Network)
Linear-evaluation-based approach
둘의 장점만 활용하기 위해, training phase에는 Meta learning을 피하고 (빠른 수렴을 위해) test phase에는 linear evaluation을 피하고 싶음. -> testing phase에는 prototype classifier를 쓰는 데에 집중, training은 standard classification manner로 하는 방식 채택.
즉, 학습된 feature extractor에다가 prototype classifier 적용하는 것. 어떤 feature extractor에든 적용 가능. linear-evaluation methods(logistic regression/SVM)보다 간단함.
이들은 목표가, linear-evaluation methods를 prototype classifier로 대체할 것이다! 이게 아니라, prototype classifier가 linear-evaluation의 짝꿍(?)으로 사용될 수 있고, few-shot learning 문제를 다루는 실용적인 첫번째 step이 될 수 있다는 걸 보이는 것이라고 함.
NeurIPS 2022 paper : https://openreview.net/pdf?id=U_hOegGGglw