Segmented Overdetermined Nonlinear Independent Component Analysis for Online Neural Filtering

In particle collider experiments a huge amount of information is produced, but only a small part is relevant for physics characterization. An efficient filtering (trigger) system is required to guarantee that valuable signatures will be recorded and most of the background noise rejected. In previous works the standard linear independent component analysis (ICA) model was used for feature extraction and good results were obtained, but it is known that the measured signals are modified by some sort of nonlinear phenomena. Another characteristic of our particular application is that there exists more sensors than original sources (the problem is over determined). In this work is proposed a novel structural model for the over determined Non-linear ICA problem. Multi-layer perceptron networks were applied for nonlinear mixing function estimation. The extracted nonlinear independent components were used to feed a neural filter that performs online particle classification with high discrimination performance.