What is the correct architecture for a time series predicting LSTM neural network?

Question

Concerning the structure of a LSTM network

If I wanted to create LSTM network for solving time series predictions, how should I structure the hiddens layers of the neural network?

• A LSTM memory block would represent a hidden layer and all the nodes in the layer would be represented by cells?
• Each hidden layer should consist of numerous LSTM memory blocks, and a collection of such blocks will form a layer?

---

Graphical representation:

Either in this manner:

Or like this ?

Answer 1

After talking to some of the professors at my university, I finally got this sorted out.

You should view a LSTM block as a single neuron in your network.

Thus this network would be regarded as a neural network with a single hidden layer with two neurons:

Answer 2

As far as I know you'd need multiple LSTMs to form a layer (see this paper for example) but you should probably ask this on https://stats.stackexchange.com/