question about LSTM states about dtln HOT 3 CLOSED

Hi,

for a good performance you need the states. Without states you don't have any modelling/memory over time. You can try to implement the LSTM as multiple fully connected layers, but you need support of a concatenate operation and still have to pass the states of the last time step to the current one.

from dtln.

Hi,
I'm confused about You can try to implement the LSTM as multiple fully connected layers. Do you mean I should replace LSTM with multiple fully connected layers and then concatenate the states of the last time step to the current input as a total input?

Looking forward to your reply.

from dtln.

An LSTM is basically one concatenation, four matrix multiplications and three dot products. So it can be implemented with fully connected layers, dot product operation and concatenate operation, when just one time step is calculated. For more on the theory of LSTMs see this wonderful blog post: https://colah.github.io/posts/2015-08-Understanding-LSTMs/

from dtln.