ProtoPNet proposed by Chen et al.is able to provide interpretability that conforms to human intuition,but it requiresmany iterations of training to learn class-specific prototypes and does not support few-shot learnin...ProtoPNet proposed by Chen et al.is able to provide interpretability that conforms to human intuition,but it requiresmany iterations of training to learn class-specific prototypes and does not support few-shot learning.We propose the few-shot learning version of ProtoPNet by using MAML,enabling it to converge quickly on different classification tasks.We test our model on the Omniglot and MiniImagenet datasets and evaluate their prototype interpretability.Our experiments showthatMAML-ProtoPNet is a transparent model that can achieve or even exceed the baseline accuracy,and its prototype can learn class-specific features,which are consistent with our human recognition.展开更多
文摘ProtoPNet proposed by Chen et al.is able to provide interpretability that conforms to human intuition,but it requiresmany iterations of training to learn class-specific prototypes and does not support few-shot learning.We propose the few-shot learning version of ProtoPNet by using MAML,enabling it to converge quickly on different classification tasks.We test our model on the Omniglot and MiniImagenet datasets and evaluate their prototype interpretability.Our experiments showthatMAML-ProtoPNet is a transparent model that can achieve or even exceed the baseline accuracy,and its prototype can learn class-specific features,which are consistent with our human recognition.
