• Slides
• Our lecture, seminar
• The only relevant video-lecture we could find - video
• Will hopefully record our lecture in english soon!
• Self-critical sequence traning original article

• An awesome post explaining attention and long-term memory models.
• BLEU and CIDEr articles.
• Image captioning
  • MSCOCO captioning challenge
  • Captioning baseline notebook
• Other articles on reinforcement learning for natural language:
  • task-oriented conversation system
  • generating dialogues
  • sequential adversarial networks (a.k.a. SeqGAN)
  • A large overview for machine translation (touching on RL, including RL failures) - article
  • How not to evaluate conversation models - article
• Overview of other non-games applications ("that article again") - https://arxiv.org/abs/1701.07274

As usual, go to practice_theano.ipynb or practice_tf.ipynb and follow instructions from there.

Other frameworks: as usual, your task remains the same as in the main track:

• Implement or borrow seq2seq model for the same translation task
  • Neat tenworflow repo
  • Important - this repo uses simplified phoneme dict - make sure you change preprocessing phase to meaningfully compare results.
• Implement self-critical sequence training ( = basic policy gradient with a special baseline, see notebook)
• Beat the baseline (main notebook: step6)

Even if you decide to use custom frameworks, it is highly recommended that you reuse evaluation code (e.g. min Levenshtein) from the main notebook to avoid confusion.