We have developped an unconventional solution to the problem of finding the vertex of high energy interactions recorded in a collider experiment. The solution is based on two feed-forward neural networks with fixed architectures. The system, tested on simulated data sets, was shown to perform better than conventional
algorithms. Here we report on further improvements, which lead to a substantial shrinkage of the size of the network. We also show how the parameters of the network can be changed so as to build an efficient pattern recognition tool for tracks originating from the vertex.