Beginning Anomaly Detection Using

Lighter on memory: LSTMs store information in their cell gates so 

if the input sequence is long, much more memory is used by the 

LSTM network. Comparatively, TCNs are relatively straightforward 

because they are comprised of several layers that all share their own 

respective filters. Compared to LSTMs, TCNs are much lighter to run 

in regards to their memory usage.

However, TCNs do carry some disadvantages:

• 

Memory usage during evaluation mode: RNNs only need to 

know some input xt to generate a prediction, since they maintain 

a summary of everything they learned through their hidden state 

vectors. In comparison, TCNs need the entire sequence up until the 

current point again to make an evaluation, leading to potentially 

higher memory usage than an RNN.

Chapter 7   temporal Convolutional networks

---

259

•