An electronic system for simulation of neural networks with a micro- second real time constraint

Neural networks implemented in hardware can perform pattern recognition very quickly, and as such have been used to advantage in the triggering systems of certain high energy physics experiments. Typically, time constants of the order of a few microseconds are required. We present a new system, MAHARADJA, for evaluating MLP and RBF neural network paradigms in real time. The system is tested on a possible ATLAS muon triggering application suggested by the Tel Aviv ATLAS group, consisting of a 4-8-8-4 MLP which must be evaluated in 10 microseconds. The inputs to the net are dx/dz, x(z=0), dy/dz, and y(z=0), whereas the outputs give pt, tan(phi), sin(theta), and q, the charge. With a 10 MHz clock, MAHARADJA calculates the result in 6.8 microseconds; at 20 MHz, which is readily attainable, this would be reduced to only 3.4 microseconds. The system can also handle RBF networks with 3 different distance metrics (Euclidean, Manhattan and Mahalanobis), and can simulate any MLP of 10 hidden layers or less. The electronic implementation is with FPGAs, which can be optimized for a specific neural network because the number of processing elements can be modified. (3 refs).