Federated Dictionary Learning from Non-IID Data

The problem of learning a common dictionary by following a federated learning framework, in a network where each edge user may have statistically different data, is considered. In such a challenging setting, the Federated Averaging solution is shown to exhibit poor performance. To alleviate the drawbacks of this approach, two more elaborate schemes are proposed. The new schemes are designed so as to offer increased performance in the case of non-i.i.d. data, that is the focus in this work. Extensive simulation results are presented for an image processing scenario, that demonstrate the effectiveness of the proposed methods.