I have a poster to prepare
I 2pv* a Uoin8r to prTpa@e
POS Tagging
Protein folding
Spelling correction
Betty Fabre
Ph.D student
Tanguy
Urvoy
Orange Labs
Damien
Lolive
Jonathan
Chevelu
HPC Summer School , July2019
Encoder
Decoder
Input features x
Structured output
y = [START, y1, y2, .., yn, STOP]
Fixed length representation of x
ref : [2014] Sequence to Sequence Learning with Neural Networks, Ilya Sutskever, Oriol Vinyals, Quoc V. Le
Structured
output y
Â
START
Å·1
Å·2
...
STOP
Model
Model
Model
ref : [2014] Sequence to Sequence Learning with Neural Networks, Ilya Sutskever, Oriol Vinyals, Quoc V. Le
START
Å·1
Model
START
Å·1
Å·2
...
STOP
Model
Model
Model
START
Å·1
Model
START
Å·1
Å·2
...
STOP
Model
Model
Model
START
Å·1
Model
START
Å·1
Å·2
...
STOP
Model
Model
Model
y1
yt
START
Å·1
Model
START
Å·1
max p(yt|y1, y2, ..yt-1)
...
STOP
Model
Model
Model
y1
yt
START
Å·1
Model
START
Å·1
Å·t =Â argmaxy' p(y'|Å·1,..Å·t-1)
...
Model
Model
→ decoding local while
evaluation is global
→ exposure bias       Â
REINFORCE Algorithm
training with exposure bias
training with expectation
Actor-Critic Algorithm
ref:Â
- [1992] Simple statistical gradient-following algorithms for connectionist reinforcement learning, Ronald J. Williams
- [2015] Sequence Level Training with Recurrent Neural Networks, Marc'Aurelio Ranzato, Sumit Chopra, Michael Auli, Wojciech Zaremba
- [2017] An Actor-Critic Algorithm for Sequence Prediction, Dzmitry Bahdanau, Philemon Brakel, Kelvin Xu, Anirudh Goyal, Ryan Lowe, Â Â Â Joelle Pineau, Aaron Courville, Yoshua Bengio
Â