Martin
& Cundy (2018) focus on parallelizing linear RNNs and propose the
GILR (Generalized Linear RNN) architecture. GILR is used as a linear
surrogate for the hidden state dependencies of traditional LSTMs,
allowing for parallelization. The resulting architecture GILR-LSTM
retains much of the complexity of LSTMs but with parallelizability,
resulting in a larger memory footprint due to the use of surrogate
states.
Martin & Cundy (2018) focus on parallelizing linear RNNs and
propose the GILR (Generalized Linear RNN) architecture. GILR is used as a
linear surrogate for the hidden state dependencies of traditional
LSTMs, allowing for parallelization. The resulting architecture
GILR-LSTM retains much of the complexity of LSTMs but with
parallelizability, resulting in a larger memory footprint due to the use
of surrogate states.
In contrast, our work takes a different approach by simplifying traditional RNN architectures