What is the activation in an LSTM and fully connected layer?

In the documentation, it is not clear what is the activation after an lstm or a fully connected layer. In an example the structure of the network was the following: -Sequence input -LSTM layer -LSTM layer -Fully Connected Layer -Regression Layer

Someone had a similar question and the verified answer was that the activations can be imported as individual layers (e.g reluLayer) but in the example above there are no reluLayers or something similar which means that the activations must be already inside the layers (e.g inside the LSTM layer). Could someone tell me what are those activations and if it is possible to change them?

[Christos Chrysafis]

I am using my gpu to train the network and I have seen that cudnn is used in that case and the activation used in cudnn files for lstm is tanh.

Answer 1

[Astarag Chattopadhyay]

Hi Christos,

Long Short-Term Memory networks have tanh and sigmoid as the internal activation functions. You can see more details on that in the following documentation page:

https://in.mathworks.com/help/nnet/ug/long-short-term-memory-networks.html