Abstract

Complex numbers have long been favoured for digital signal processing, yetcomplex representations rarely appear in deep learning architectures. RNNs,widely used to process time series and sequence information, could greatlybenefit from complex representations. We present a novel complex gate recurrentcell. When used together with norm-preserving state transition matrices, ourcomplex gated RNN exhibits excellent stability and convergence properties. Wedemonstrate competitive performance of our complex gated RNN on the syntheticmemory and adding task, as well as on the real-world task of human motionprediction.

 

Quick Read (beta)