Learning photonic neural network initialization for noise-aware end-to-end fiber transmission

Deep Learning (DL) has dominated a wide range of applications due to its state-of-the-art performance. Novel approaches introduce Artificial Neural Networks (ANNs) on fiber communication channels to be used as intensity modulation/direct detection (IM/DD) systems and optimized in an end-to-end fashion. Despite the potential of these methods, the demanding nature of DL models limits their applications in such domains, where fast inference and low power consumption is required. Indeed, these limitations fueled the research on neuromophic architectures, including neuromorphic photonics, which holds the credentials for unlocking matrix multiplications at high frequencies, while minimizing energy consumption. However, at the same time, photonic architectures impose new challenges to DL training due to the underlying hardware constraints. In this paper, we present a trainable data-driven noise-aware initialization method oriented to easily saturated activation functions, such as those typically used in optical neurons on the transmitter and receiver side of a noisy IM/DD system intercepted by a noisy channel. The proposed method is evaluated on a fully optical IM/DD system using different fiber lengths, overcoming issues such vanishing gradient phenomena that profoundly hinders the training of the receiver and transmitter photonic neural networks (PNNs), while also improving robustness to noise.