Data Science and Machine Learning
Session Language |English
I will review the neural network techniques for signal recovery in optical transmission systems, where the neural network equalizers have to deal with considerable nonlinearity and with bi-directional memory effects. Moreover, the problems are characterized by the presence of essential noise. I will concentrate on the specific requirements that we have to impose on the proprieties of efficient neural network equalizers in optical transmission, in particular, the ultra-high accuracy (99.999%), which makes the application of some existing techniques from image recognition and computer vision rather impractical, and the requirement for the reduced complexity of the neural architectures. We will overview the problems that arise when applying neural networks in such nonstationary nonlinear signal processing tasks (high overfitting tendencies and the risk of overestimation), and address the peculiarities of pruning and quantization techniques that arise in optical signal processing.