Generative RNN

• initial sequence is used as seed and output is sampled 
  • random or argmax to sample
  • normally not taking argmax but sample with respective Softmax probabilities → allows to generate something different than input
• new output is used as seed to generate next 
• repeat until termination criterion