Assignment 3 Checklist

Thresholds

• 3 or 5 epoch burn-in
• Dev loss has either gone down less than some epsilon or has increased relative to last epoch

Experiments with Czech

• Average and Max embedding combinations
• Try different thresholds/burn-in combinations

Autoencoder

• Change embedding combination to retrieve all sets that it needs to combine
• Pass sets into separate module which contains RNN
• Split sets into train/dev
• Train RNN
• Compute embeddings once it gets below certain loss on dev
• Retain RNN, but discard optimizer.
• Repeat on future runs:
  • Should we retrain? Should we only evaluate the trained RNN on the new examples? Should we retain the old training data (previous combinations), or only use the new training data?