Multi-Resolution Siamese Networks for One-Shot Learning

Abstract

Few-shot learning, a rapidly evolving theme in deep learning research, aims to endow artificial intelligence with the same ability of humans to assimilate new information very quickly. Siamese networks have been used in this context to learn similarity between image pairs and quickly classify novel objects. This work proposes an improved architecture and a novel training method that increases a 1-shot 5-way classification accuracy on 5 entirely novel classes by around 5%, 19%, 18% and 13% respectively compared to vanilla Siamese networks when tested on Omniglot, Tiny-Imagenet, CIFAR100 as well as a custom dataset recorded with an event-driven camera. These networks, when run on a Jetson TX2 GPU can be executed within 108 ms on average.