Attention Based Siamese Networks for Few-Shot Learning
Junhua Wang, Zijiang Zhu, Jianjun Li, Junshan Li · 2018
We propose Attention based Siamese Networks for the problem of few-shot classification, where a classifier must generalize to new classes not seen in the training set, given only a few examples of each new class. In this work, we construct an efficient convolution networks to learn embedding function, and measure the similarity of two feature vectors with simple attention kernel function. Compared with state-of-the-art methods, the proposed method provides improved performance one-shot learning. Our experimental results on dataset of Omniglot and miniImageNet validate the effectiveness of the proposed method.