Open hi-bingo opened 4 years ago
why do they need multiple encoder-decoders in Fig 4? what are their purpose? In terms of skip connection, how about randomly choosing a node, and skipping that node, instead of fix length skip. So such randomly choosing skipping will look like drop-out technique. Such random skipping may be good for robustness.
Paper
Outlier Detection for Time Series with Recurrent Autoencoder Ensemble IJCAI'19
Code
https://github.com/tungk/OED
Authors
Task
Outlier Detection for Time Series
Datasets
RNN ensemble framework
Consider Recurrent Skip Connection Networks (RSCNs) (ref Wang and Tian, 2016). Set different L for skip connection and the function:
For ensemble framework, we need different structures. So we random remove some connection. Specifically, we introduce a sparseness weight vector
, and make sure each
at least have one element equal to 1.
The total sparsely-connected RNNs (S-RNNs) can be expressed as
And the figure shows the final structure
![SRNN](https://user-images.githubusercontent.com/14290851/63739048-7859c000-c8be-11e9-8dfb-2b5929e9269c.png)
Independent Framework
Set different w_t for each autoencoder, and train the autoencoders independently.![image](https://user-images.githubusercontent.com/14290851/63746622-d941c200-c8d7-11e9-80a8-df5bc28692ee.png)
Shared Framework
Add a shared layer and concatenate all encoders' codes and use it as init hidden state for all decoder.![image](https://user-images.githubusercontent.com/14290851/63746687-00988f00-c8d8-11e9-813d-745a054cac1b.png)
Questions
Does anyone know more about different skip connection strategies in RNN?