Skip Residual Pairwise Networks With Learnable Comparative Functions for Few-Shot Learning

摘要

In this work we consider the ubiquitous Siamese network architecture and hypothesize that having an end-to-end learnable comparative function instead of an arbitrarily fixed one used commonly in practice (such as dot product) would allow the network to learn a final representation more suited to the task at hand and generalize better with very small quantities of data. Based on this we propose Skip Residual Pairwise Networks (SRPN) for few-shot learning based on residual Siamese networks. We validate our hypothesis by evaluating the proposed model for few-shot learning on Omniglot and mini-Imagenet datasets. Our model outperforms the residual Siamese design of equal depth and parameters. We also show that our model is competitive with state-of-the-art meta-learning based methods for few-shot learning on the challenging mini-Imagenet dataset whilst being a much simpler design, obtaining 54.4% accuracy on the five-way few-shot learning task with only a single example per class and over 70% accuracy with five examples per class. We further observe that the network weights in our model are much smaller compared to an equivalent residual Siamese Network under similar regularization, thus validating our hypothesis that our model design allows for better generalization. We also observe that our asymmetric, non-metric SRPN design automatically learns to approximate natural metric learning priors such as a symmetry and the triangle inequality.