Statistical Wide-Sense Curriculum Learning for Neural Network-Based Pre-Distorter in Coherent Optical Transmitters

Optical communications systems' performance is limited by physical layer nonlinearities inherent in some of their transmitter components. In order to overcome some of these limitations, several approaches utilizing artificial neural networks to digitally pre-distort the transmitted signal have been proposed in the last few years, alongside other more classical models. However, their performance was strongly dependent on the specific training sequence used. This work improves the performance of neural network-based direct learning approach for digital pre-distortion applications. The proposed method, which is based on the curriculum learning approach, aims to construct a training paradigm of gradually increasing complexity based on the statistical properties of the input signal, in order to achieve better convergence of the network. We prove that the proposed method defines a proper curriculum regardless of the training data distribution, considering the proposed wide-sense curriculum conditions and the case of binary weighting functions. A comparative analysis examines a pre-distorter for a coherent optical transmitter modelled by a Long Short-Term Memory neural network. An improvement of 4.8dB in terms of normalized mean-square error is demonstrated for the case of a 3-bits linear quantizer, compared to the conventional training scheme.