Meta-Learning with Memory Augmented Neural Networks

A chainer implementation of Meta-Learning with Memory Augmented Neural Networks
(This paper is also known as One-shot Learning with Memory Augmented Neural Networks )

• Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, Timothy Lillicrap, Meta-Learning with Memory-Augmented Neural Networks, [link]

• Some code is taken from tristandeleu's implementation with Lasagne.

How to run

1. Download the Omniglot dataset and place it in the data/ folder.

2. Run the scripts in data/omniglot to prepare dataset.

3. Run scripts/train_omniglot.py (Use gpu option if needed)

Summary of the paper

The authors attack the problem of one-shot learning by the approach of meta-learning. They propose Memory Augmented Neural Network, which is a variant of Neural Turing Machine, and train it to learn "how to memorize unseen characters." After the training, the model can learn unseen characters in a few shot.