Analog neural network implementations using active, distributed charge injection techniques

Implementation and circuit design issues related to active, charge-based neural network cells are discussed. Motivation for the implementation of neural systems in analog VLSI hardware using MOS charge modulation techniques is presented, and a specific circuit design for charge-injection synapses is discussed. Capacitive weight storage and charge updating are discussed. The overall architecture for a two-layer feedforward network using these techniques is presented. Simulation and experimental results indicate that charge-based analog neural computation schemes may offer significant advantages over classical static analog or digital implementations of neural systems.<
>