The non-infinite contrast of the spatial light modulators used to modulate the recording and readout beams in dynamic holographic interconnects induces noise. We show that this noise can be easily computed. It increases with the system capacity. If nothing is done to prevent it, then most of the holographic neural networks previously proposed in the literature are limited to uninteresting capacities. Based on the deterministic nature of this noise source, we propose various techniques to reduce it. These techniques can be easily implemented on any setup. We believe that the resulting gain in capacity considerably renews the interest for these holographic neural networks.