L.P. Caloba, H. Gottschalk and J.M. Seixas
EE/COPPE/UFRJ
Rio de Janeiro, Brazil
AbstractA high-performance electron/pion discriminator has been designed using neural processing on scintillating fiber calorimeter data [1]. For this design, neural networks using hyperbolic tangent as the activation function are easily trained by using backpropagation, but the implementation of the hyperbolic tangent on integrated circuits presents difficulties in terms of precision and speed. On the other hand, networks using threshold neurons can achieve quite fast processing times and are easily implemented for high precision performance, but multilayer networks of this kind present difficulties in wich training is concerned.
A new training procedure for neural networks is developed in this paper in order to achieve a very fast electron/pion discrimination by means of analogue integrated circuit implementation. Departing with hyperbolic tangent neurons, the training procedure, in the first phase, makes the network to learn the particle discrimination while efficiently saturates each output neuron (including those at the hidden layer). Then, the activation function is switched to a signal function, so that the network acts as a multilayer perceptron. As a consequence, the neural discriminator can be implemented using fast comparators and resistive networks, which are realized by active elements in the final design. This simple and sufficiently precise analogue implementation can achieve processing times of the order of a few nanoseconds.
The described technique is applied to a fiber calorimeter, using experimental data obtained from testbeam. For 98% electron efficiency, the neuron solution achieves pion rejection factors of up to several hundred in the energy range of 40 to 150 GeV and at different impact points, which improves the performance of classical methods.
Reference: