Extrapolation of Periodic Signal with Poisson Noise Using Neural Networks

Designing an algorithm capable of extrapolating a periodic signal with Poisson noise turns out to be a difficult problem. In this work, we focused on the extrapolation of this kind of a signal using a convolutional neural network and recurrent neural network. It appears that simulations with real noise level is problematic. For extrapolation to be correct, we need an algorithm that can either predict the signal with real Poisson noise or the pure signal without the noise (Poisson noise can be subsequently added). This means that it must find the character of the signal (period, amplitude, phase, and modulation of the signal). It turns out that convolutional and recurrent neural networks are capable of just such an extrapolation. Next, we focused on the methods of data preprocessing (standardization and normalization) and found out how they affect the results. Although preprocessing is used to make machine learning algorithms work faster and more accurately, it turns out that it is not always the case.