Hardware-oriented learning for cellular neural networks

This paper presents an approach to learning, which focusses on finding a set of parameter values taking into account the nonidealities of a specific implementation. Therefore learning is done on a more accurate model of a CMOS cell, and not on the original CNN model proposed in 1988. This hardware-oriented approach will be applied to a current-mode CNN-model based on the full-signal-range model published in [10, 2], where the dynamic block consists of two current mirrors. It is shown, that a two-quadrant multiplier is sufficient for the multiplication with the template coefficients, by changing the model, further reducing the area consumption. Using a hardware-oriented approach to learning thus not only allows to find template values for a specific VLSI-implementation, but may also lead to further simplifications of CNN-implementations.