We have investigated the problem of track finding with a recurrent neural network algorithm based on the Hopfield model and considered the possibility of a hardware implementation with DSP’s. Starting from a set of signal points we define track segments and set a cut on the length to keep the size of the network reasonable. Those segments surviving the cut are associated to neurons. A geometric coupling of neighbouring segments is used to select smooth combinations of them. Given random initial conditions the network converges to a solution. The method may be applied to a variety of curves.
Dynamic Approximation with Feedback Control for Energy-Efficient Recurrent Neural Network Hardware
