junxnone / aiwiki

AI Wiki
https://junxnone.github.io/aiwiki
18 stars 2 forks source link

ML AutoML NAS NASwRL #390

Open junxnone opened 4 years ago

junxnone commented 4 years ago

Neural Architecture Search with Reinforcement Learning

Pipeline

image

without skip-connections + skip-connections
image image

Training

Name Formula
Reward image
REINFORCE image
empirical approximation image
with Baseline Function image
Parallelism and Asynchronous Updates

image

Skip-Connections

当前RNN 状态和前 N-1 Anchor Point 函数 image
hj RNN 第 j 层 Anchor Point 隐藏状态
j 0 ~ N-1

RNN Cell

RNN Cell 1 - Base Cell image
RNN Cell 2 - LSTM image
Input Base Cell - xt/ht-1
LSTM - xt/ht-1/ct-1
Output Base Cell - ht
LSTM - ht/ct
Combination method Addition/Elementwise/Multiplication
Activation Function tanh/sigmoid

Reference

junxnone commented 4 years ago

junxnone/tech-io#805 junxnone/aiwiki#395