Non-linear circuit effects on analog VLSI neural network implementations
Miguel Onorato, Maurizio Valle, Daniele D. Caviglia, Giovanni Maria Bisio · 2002
We present an analog VLSI neural network for texture analysis; in particular we show that the filtering block, which is the most critical block of the architecture for precision of computation, can be implemented using simple and compact analog circuits, without significant loss in classification performance. Through an accurate analysis of the circuits it is possible to model the real circuit characteristics in the software simulation environment; the weights calculated in the learning phase (which is performed off-line using the adaptive simulated annealing algorithm), can be properly coded into analog circuit variables in order to implement the correct operation of the network.