Poster Title:  Structured Prediction with autoregressive models and reinforcement learning
Poster Abstract: 

Structured Prediction has received a renewed interest from the deep learning community since the introduction of encoder-decoder models such as Seq2Seq and Transformers and attention mechanisms.
This presentation reviews the use of autoregressive neural network architectures for structured prediction with a focus on text generation tasks. I will also discuss how reinforcement learning algorithms such as REINFORCE or Actor-Critic can be implemented to train such architectures and how it offers a solution to the well-known exposure bias problem.

keywords: structured prediction, deep learning, encoder-decoder, text generation, reinforcement learning
Poster ID:  D-1
Poster File:  HTML document HPCSS_19.html
Poster Image: 
Poster URL: 
 
Show comments