Mixed-Mode Analog VLSI Continuous-Time Recurrent Neural Network

B.E. Brown, X. Yu, and S.L. Garverick (USA)


analog VLSI, CTRNN, mixed-mode, neural networks,hardware neural networks.


Signal processing circuits for a Continuous-Time Recurrent Neural Network have been implemented using subthreshold analog VLSI. In this mixed-mode (current and voltage) approach, state variables are represented by voltages while neural signals are conveyed as currents, distinguishing the present work from previous approaches that employ analog VLSI technology. The use of current allows for the accuracy of the neural signals to be maintained over long distances, making this architecture highly robust and scalable. The mixed-mode network architecture is described, and design details and test results for synapse, sigmoid and fully interconnected two-neuron oscillator circuits are given. The synapse and neuron circuits are based on unique implementations of the subthreshold Gilbert-core multiplier and employ just three and seven transistors per cell, respectively, and both have measured accuracy better than 1%. A fully interconnected, 2-neuron/4-synapse array was programmed to perform a simple oscillator function, and measured results closely match software simulation using both SPICE and a high-level C++ model. The synapse cell is just 43.2 x 42.6 m2 in a 1.2-m CMOS process., so much larger arrays are possible.

Important Links:

Go Back