Learning to Learn in Chainer

A chainer implementation of "Learning to learn by gradient descent by gradient descent" by Andrychowicz et al. It trains and tests an LSTM-based optimizer which has learnable parameters transforming a series of gradients to an update value.

What is LSTM-based optimizer?

SGD is a simplest transformation of gradient; just multiplying a gradient with a constant value of learning rate. Momentume SGD or more sophiscated ones like Adam use a series of graadients. Finally, this LSTM-based optimizer is a general case of them. While they define update formula in advance (by hand), the LSTM-based optimizer learns the formula, i.e, how to merge a history of gradients to reach convergence efficiently.

Dependencies

See requirements.txt

• chainer>=6.0.0
• cupy>=6.0.0
• seaborn==0.9.0

Training

This repository includes an experiment using a fully connected network on MNIST.

python train_mnist.py -g 0

The script trains an LSTM-based optimizer as follows

init optimizer
for-loop
  init model
  for-loop
    update model by optimizer
    update optimizer by Adam (at every 20 steps)
  test optimizer (through another training of models)
  save optimizer if it is best

Citation

@incollection{NIPS2016_6461,
  title = {Learning to learn by gradient descent by gradient descent},
  author = {Andrychowicz, Marcin and Denil, Misha and G\'{o}mez, Sergio and Hoffman, Matthew W and Pfau, David and Schaul, Tom and Shillingford, Brendan and de Freitas, Nando},
  booktitle = {Advances in Neural Information Processing Systems 29},
  year = {2016},
  url = {http://papers.nips.cc/paper/6461-learning-to-learn-by-gradient-descent-by-gradient-descent.pdf}
}