Paper
https://www.isca-archive.org/interspeech_2014/sak14_interspeech.pdf
Equations
Modifies the PeepholeLSTM with the following addition
$$\begin{align}
\mathbf{i}_t &= \sigma\left( \mathbf{W}_{ix} \mathbf{x}_t + \mathbf{W}_{ih} \mathbf{h}_{t-1} + \mathbf{W}_{ic} \mathbf{c}_{t-1} + \mathbf{b}_i \right), \\\
\mathbf{f}_t &= \sigma\left( \mathbf{W}_{fx} \mathbf{x}_t + \mathbf{W}_{fh} \mathbf{h}_{t-1} + \mathbf{W}_{fc} \mathbf{c}_{t-1} + \mathbf{b}_f \right), \\\
\mathbf{c}_t &= \mathbf{f}_t \circ \mathbf{c}_{t-1} + \mathbf{i}_t \circ g\left( \mathbf{W}_{cx} \mathbf{x}_t + \mathbf{W}_{ch} \mathbf{h}_{t-1} + \mathbf{b}_c \right), \\\
\mathbf{o}_t &= \sigma\left( \mathbf{W}_{ox} \mathbf{x}_t + \mathbf{W}_{oh} \mathbf{h}_{t-1} + \mathbf{W}_{oc} \mathbf{c}_t + \mathbf{b}_o \right), \\\
\mathbf{h}_t &= \mathbf{o}_t \circ h(\mathbf{c}_t), \\\
\mathbf{r}_t &= \mathbf{W}_{rh} \mathbf{h}_t, \\\
\mathbf{y}_t &= \phi(\mathbf{W}_{yr} \mathbf{r}_t + \mathbf{b}_y).
\end{align}$$
Since this is separate from the actual equations we can also make this a generic wrapper layer. We have to provide
- Cell as implemented in the paper
LSTMPCell
- Additional wrapper
RecurrentProjection() that takes a cell and returns
$$\begin{align}
\mathbf{r}_t &= \mathbf{W}_{rh} \mathbf{h}_t, \\\
\mathbf{y}_t &= \phi(\mathbf{W}_{yr} \mathbf{r}_t + \mathbf{b}_y).
\end{align}$$
We also have to find a way to stack it I suppose.
Official implementation
none
Paper
https://www.isca-archive.org/interspeech_2014/sak14_interspeech.pdf
Equations
Modifies the PeepholeLSTM with the following addition
Since this is separate from the actual equations we can also make this a generic wrapper layer. We have to provide
LSTMPCellRecurrentProjection()that takes a cell and returnsWe also have to find a way to stack it I suppose.
Official implementation
none