(NIPS 2016) Matching Networks for One Shot Learning

Keyword [Bi-LSTM] [Matching Net] [Attention LSTM]

Vinyals O, Blundell C, Lillicrap T, et al. Matching networks for one shot learning[C]//Advances in neural information processing systems. 2016: 3630-3638.

1. Overview

In this paper, it proposes Matching Nets (MN)

  • maps a small labeled support set and an unlabeled example to its label
  • based on Bi-LSTM and Atten-LSTM
  • experiments on vision (Omniglot, ImageNet) and language (Penn Treebank)
  • devise new data set. miniImageNet (60,000 images, 84x84, 100 classes, 600 examples/class)

1.1. Model

1.1.1. Basic Formulation

  • x_i, y_i. labeled small support set
  • a. attention mechanism

1.1.2. Attention Kernel

  • g, f. embedded function
  • c. cosine similarly
  • x^. query image

1.1.3. Full Context Embeddings

  • bi-LSTM

  • S should be able to modify how to embed the test img x^ through f

  • K. number of unrolling steps of LSTM

1.2. Loss Function

1.2.1. Training

  • for each episode, sample L labels
  • sample support set S and query set B

1.2.2. Testing

  • know support set S’
  • predict label of query set B’

1.3. Network

  • 4 Conv
  • [64, 3x3 Conv; BN; ReLU; Maxpool]

1.4. Experiments

  • Omniglot

  • ImageNet