aicoe-kaggle / diabetic-retinopathy

Other
0 stars 0 forks source link

Neural Architecture Search #9

Open TreeinRandomForest opened 3 years ago

TreeinRandomForest commented 3 years ago

Do this with MNIST (or a simple dataset) before the retinopathy data.

Goal: to learn an architecture that gives maximal validation score

Techniques:

  1. Use the NAS module in dragonfly (https://github.com/dragonfly/dragonfly).
  2. Implement the policy gradient (REINFORCE) algorithm (used in the cartpole issue) to learn an architecture. See: https://arxiv.org/abs/1611.01578 and also, https://arxiv.org/abs/1802.03268 for a much more efficient approach.

Make the following plot: x-axis: train time y-axis: validation score

Inspect the architectures learned. Use a small set of primitives (conv layers, pool layers, a few fixed kernel sizes etc.)