Résumé: In this paper, we propose two algorithms based on the random vector functional link network (RVFLN) and alternating direction method of multipliers algorithm to solve distributed learning (DL) problems with datasets containing outliers. In distributed scenarios, training datasets are separately divided and stored on each node of the communication network and cannot be shared or collected over the communication network due to privacy-preserving policy and network environmental restrictions. Thus, each node of the communication network only deals with local data and shares updated output weights of the global RVFLN model with its neighbors. However, the majority of existing DL algorithms are not robust enough when the dataset contain outliers. To overcome the this drawback, we intend to apply the norm and Huber based error terms to the global loss function and propose the corresponding distributed robust learning algorithms and denote them as the L1-DRVFL and Huber-DRVFL algorithms, respectively. The proposed algorithms work in fully distributed manner and are privacy-preserving methods. Experiments show that the proposed algorithms are robust and efficient in distributed learning with data including outliers. Moreover, the Huber-DRVFL algorithm is more stable than the L1-DRVFL algorithm when the parameters vary.

Langue: Anglais