Hi,
The example mentioned in Sequence-to-Sequence Regression Using Deep Learning uses a multi-layer network that is why a fullyConnectedLayer in between is connected. You may also try the example without it but the number of parameters in the Network will decrease which may affect the overall performance (in terms of accuracy) of the Network. The last layer is fullyConnectedLayer(numResponse) which is according the guideline that are mentioned in here.
It’s a general practice in Deep Learning to include fullyConnectedLayer with appropriate outputSize to make the network deep and using dropoutLayer to prevent the network from overfitting.
