ANALOG DENDRO-DENDRITIC ARRAYS WITH DIGITAL ON-CHIP LEARNING
A design of arrays of CMOS analog Dendro-dendritic Artificial Neural Network (DANN) chips with on-chip digital learning is described. The building blocks, namely, the neuron unit and the adaptive synapse, are employed to construct several architectures. One design comprises a reconfigurable fully-connected array chip integrating 50-neurons. A second array chip design integrates (9×9=81) neuron units and 825 locally-connected reconfigurable weights. In all cases, a connection is realized as a single (nonlinear) transistor with adaptive digital circuitry. The chip is designed and fabricated in 6.8 mm× 4.6 mm chip size using 2 μm CMOS technology. As an example of an application of the fabricated dendro-dendritic neural chips, real-time experiments are described in which the chips are used as a parallel digital coprocessor to demonstrate their applicability as pattern associators. These experiments entail learning an arbitrary binary image in about 10 ns with guaranteed learning capability. The stored image can subsequently be retrieved by images distorted by binary-noise in the order of 100 ns. The power dissipation of these chips in steady state is less than 5 mW using 0/5 V power supply.