Analog VLSI neural chips for real-time identification and control

Analog CMOS neural chips with on-chip learning are explored to provide efficient and inexpensive electronics for various tasks in real-time identification and control. Hardware learning circuits can successfully obtain a set of synaptic weights for multilayer feedforward neural networks that approximately satisfy any nonlinear mapping in the order of milliseconds. The fast on-chip learning can identify nonlinear dynamical systems to avoid the modeling uncertainty and parameter variations in real-time. Model reference adaptive control with online identification using the neural chips are also proposed.