Lucca Gamballi, Daniel G. Tiglea, Renato Candido, Magno T. M. Silva. Uma nova estratégia para o treinamento distribuído de redes neurais. Simpósio Brasileiro de Telecomunicações – SBrT’2024, 2024, Belém, Anais do SBrT 2024, 2024, pp. 1–5. (in portuguese)

Abstract

In this paper, a new strategy for combining weights of neural networks in a distributed scenario is proposed. The distributed approach has been employed in applications that take into account a great amount of data, whose privacy must be maintained. However, if one or more nodes in the topology are contaminated by noise, the performance of the global model may be deteriorated. This occurs since the contaminated local model propagates the noise effect to the global one. The proposed strategy is evaluated with different topologies using convolutional neural networks. In order to preserve privacy, only the models weights are shared, such that a network only has access to the local dataset. Image classification results show that the proposed strategy is robust to the presence of noise.

Keywords

Neural networks, distributed processing, topology, combination strategy, noise.

Downloads

• PDF