Autoencoder Model Type

This model type of Anomaly Detection is based on a neural network that has an autoencoder structure. In particular, one of the layers of an autoencoder has less neurons than the dimensionality of the network input. This layer is referred to as "bottleneck" and forces the network to learn a lower dimensional embedding. In ASCMO-DYNAMIC, recurrent neural network layers can be used in the autoencoder. All recurrent network cells available for RNNs are also available for autoencoder:

• Basic RNN Cell

• Long Short-Term Memory Cell

• Gated Recurrent Update Cell

When using recurrent neural network cells, the autoencoder is able to learn temporal dynamics and temporal dependencies of the data. In addition, ASCMO-DYNAMIC offers to use a fully connected layer (also known as dense layer or linear layer) as layer type. Compared to an autoencoder that is based on recurrent cell types, an autoencoder based on dense layers cannot take into account the temporal dynamics of the data. Its advantage is that it has less parameters and the training process is usually faster.

Due to non-linear activation functions and the ability to stack any number of layers, autoencoders can handle highly nonlinear data. The downside is that training an anomaly detection model based on neural network is much slower than training an anomaly detection model based on PCA.