Beginning Anomaly Detection Using

Parallel computations: Convolutional networks pair well with 

GPU training, particularly because the matrix-heavy calculations 

of the convolutional layers are well suited to the structure of GPUs, 

which are configured to carry out matrix calculations that are part of 

graphics processing. Because of this, TCNs can train much faster than 

RNNs.

• 

Flexibility: TCNs can change input size, filter size, increase dilation 

factors, stack more layers, etc. in order to easily be applied to various 

domains.

• 

Consistent gradients: Because TCNs are comprised of convolutional 

layers, they backpropagate differently than RNNs do, and thus all 

of the gradients are saved. RNNs have a problem called exploding 

or vanishing gradients, where sometimes the calculated gradient is 

either extremely large or extremely small, leading to the readjusted 

weight to be too extreme of a change or to be a relatively nonexistent 

change. To combat this, types of RNNs such as the LSTM, GRU, and 

HF-RNN, were developed.

•