Figure 3.

Schematic illustration of the decoder, shown for rate $\frac{1}{3}$

The input sequence y is split into the subsequences $y_0$, $y_1$ and $y_2$, which correspond to the encoded subsequences $x_0$, $x_1$ and $x_2$, respectively. $y_0$ and $y_1$, together with a prior $p_0$, are used as the input for the first decoder network $D_0(·)$. The output posterior q and $y_0$ are subsequently interleaved and serve, together with $y_2$, as the input for the second decoder $D_1(·)$. The output of the second decoder is deinterleaved by the inverse of the interleaving function, ${\pi }^{-1}$, and is used as updated prior for the first decoder. This process will be repeated until a user-defined amount of iterations is reached. The final, deinterleaved output of the second decoder will then be used to generate the output sequence $\hat{u}$ by a sigmoid activation function σ.