Evaluating the influence of loss function on performance of a neural net- work for particle 3D shape reconstruction from 2D projections

Resumo

Abstract. In recent years, Supervised Neural Networks have been used to solve different tasks due to their ability
to extract and predict patterns. This kind of algorithm has in its development a first step known as the training
phase, where it tries to find the best values for the weights and biases, making the neural net predictions as close
as possible to some expected outputs. In each training iteration, updating weights and biases is realized like an
optimization problem. When developing a Neural Network that uses Spherical Harmonics Functions (SPH) for
reconstructing the 3D shape of a particle from its 2D projections, one can choose as objective function different
parameters, e.g., the volume, radius, the SPH coefficients, etc. In this paper, we discuss the performance of a
Neural Network trained with different loss functions in a 3D shape reconstruction problem context.