Matching Networks for One Shot Learning

Posted on 10/04/2019, in Paper.
  • Overview: This paper empower metric-learning by using neural nets to learn the non-parametric structure. It also frame the one shot learning as a meta learning problem and achieved SOTA result on mini-ImageNet and Omnilot
  • Attention Kernel: The core inference logic is very similar to nearest neighbor except it is done over a space defined by learnable embedding function f and g.
  • Support set: Support set S is sometimes a training set with limited samples per class. But the bottom line is the class in the support set covers all possible class. This also means the in the testing time it can be replaced by a different support set S’.
  • Meta learning: During training we first sample label set L from training set T, and then support set S and batch B given L. The author claims this is consistent with the testing target and could work for set S' sampled for novel labels.
  • Full Context Embedding: Not totally convinced. The author believes the embedding should be done conditioning on the support set.

I found this blog post One Shot Learning and Siamese Networks in Keras – Neural Tinkering is very helpful. It provides a solid overview of one-shot problem set-up and dataset, as well as the Siamese network, which is one of the source of this paper.