Abstract

In this paper, we compare different types of Recurrent Neural Network (RNN)Encoder-Decoders in anomaly detection viewpoint. We focused on finding themodel what can learn the same data more effectively. We compared multiplemodels under the same conditions, such as the number of parameters, optimizer,and learning rate. However, the difference is whether to predict the futuresequence or restore the current sequence. We constructed the dataset withsimple vectors and used them for the experiment. Finally, we experimentallyconfirmed that the model performs better when the model restores the currentsequence, rather than predict the future sequence.

 

Quick Read (beta)