In this work, we propose a convolutional neural network (CNN) based low-complexity approach for downlink (DL) channel estimation (CE) in frequency division duplex (FDD) systems. In contrast to existing work, we use training data which solely stems from the uplink (UL) domain. This allows to learn the CNN centralized at the base station (BS). After training, the network parameters are offloaded to mobile terminals (MTs) within the coverage area of the BS. The MTs can then obtain channel state information (CSI) of the MIMO channels with the low-complexity CNN estimator. This circumvents the necessity of an infeasible amount of feedback, i.e., acquisition of training data at the user, and the offline training phase at each MT. Numerical results show that the CNN which is trained solely based on UL data performs equally well as the network trained based on DL data. Furthermore, the approach is able to outperform state-of-the-art CE algorithms.