Input Features Selection for Neural Data Analysis in Astronomical Imaging

R. Cancelliere and M. Gai (Italy)


Simulation Tools and Techniques; neural network; data modelling


The removal of chromaticity in high precision astrometric measurements is a very important challenge because chro maticity can represent a relevant source of systematic er ror; we perform this task using a feed forward neural net work and focus on the usefulness of a proper preprocess ing applied to the network parameters. We use a few sta tistical moments properly selected with a careful prepro cessing and ļ¬ltering to face the necessity of a good choice of the input parameters that encode images; they are then used as inputs to a feed forward neural network trained by backpropagation to remove chromaticity. We show that a preprocessing devoted to analyze the input-output depen dences allows to obtain the same diagnosis performance using as inputs to the neural network less parameters with respect to the diagnosis performed without preprocessing.

Important Links:

Go Back

Rotating Call For Paper Image