Popular few-shot algorithms are trained on the mini-ImageNet dataset in both Euclidean and Non-Euclidean spaces. From our experiments we find that the non-Euclidean geometry has improved the results for some tasks as shown in the table below. Furthermore, the use of non-Euclidean space can improve the model performance where there are implicit hierarchical relations in the input.
- Transductive CNAPS
- Model-Agnostic Meta Learning (MAML)
- Matching Networks
mini-ImageNet-1 has implementations for MAML and Matching Networks, whereas mini-ImageNet-2-Transductive-CNAPS consists of the Transductive CNAPS implementation, in which we obtain the highest test accuracy of 71.4 % for 5-way 5-shot image classification using our non-Euclidean model.
5-Way 5-Shot Learning on mini-ImageNet Dataset
[1] mini-ImageNet
[2] Transductive CNAPS
[3] Non-Euclidean Library
[4] Few-Shot Algorithms