Comparative study of implementing ICNNs on FPGAs

Hardware implementation of two independent component neural networks, one that maximizes the mutual-information between input and output and one that minimizes the divergence at the output, is investigated. Table lookup and computer arithmetic implementations with field-programmable gate arrays are compared. The lookup table approach is superior, in terms of both cost and performance. The network that minimizes output-divergence is less costly to implement than that which maximizes mutual information.