High speed analog filtering using feedforward neural network architectures

The design of analog filters has been a topic of research for many years, yielding a wide variety of techniques for addressing the problem. The work described here approaches this task from a neural network perspective to obtain some of the advantages of neural systems, such as a high tolerance to component imprecision and an ability to train or adapt high order structures. Investigations of linear filter networks utilizing neural-like system topologies are presented, along with accompanying training algorithms and simulation results. Designs of several network components in 2 μm CMOS are described, with simulations indicating their potential for implementing high order, self-programming analog filters at bandwidths above 70 MHz