i'm having some trouble understanding how does the basic building block of a SCINet works. In the paper the author describes the SCI-block with the following figure:


In which $\phi$, $\theta$, $\eta$ and $\rho$ are convolutional filters with the following architecture:

Convolutional filter structure

All this is described with the equations below: Odd and even equations

What I understood is that the splitted sequences go through a gated linear unit (GLU) in order to control what information is passed from the complementary sequence. This creates $F_{odd}^{s}$ and $F_{even}^{s}$ which now contain relevant features from their complementary sequence.

However, I can't seem to understand why these enhanced sequences are further projected onto two additional hidden states with filters $\rho$ and $\eta$ in order to substract or add them to their complementary sequence. Why is this step added?

If this is a common practice in neural networks i'd be glad to know it's name and if you could point me to some reference that would be great. Thanks in advance!



You must log in to answer this question.