Variational expectation maximization attention broad learning systems

Recently, broad learning system (BLS) has received much attention due to its concise network structure and strong incremental learning ability. However, as it belongs to a simple feedforward neural network, when encountering time series with sequential characteristic, it cannot effectively fit them and finish the related tasks such as wireless traffic prediction. In this paper, we propose a kind of attention mechanism-based BLS for dealing with time series while incremental learning can still be implemented. In order to learn more information from time series with relatively low computational complexity and infer the optimal number of bases in the attention mechanism, a variational form of expectation maximization attention mechanism for BLS is proposed. Experimental results show that the proposed algorithm can achieve good performance in both classification and regression tasks associated with time series.