Abstract
In scientific research and engineering practice, the design of deep spiking neural network (DSNN) architectures remains a complex task that heavily relies on the expertise and experience of professionals. These architectures often require repeated adjustments and modifications based on factors such as the DSNN's performance, resulting in significant consumption of human and hardware resources. To address these challenges, this paper proposes an innovative evolutionary membrane algorithm for optimizing DSNN architectures. This algorithm automates the construction and design of promising network models, thereby reducing reliance on manual tuning. More specifically, the architecture of DSNN is transformed into the search space of the proposed evolutionary membrane algorithm. The proposed algorithm thoroughly explores the impact of hyperparameters, such as the candidate operation blocks of DSNN, to identify optimal configurations. Additionally, an early stopping strategy is adopted in the performance evaluation phase to mitigate the time loss caused by objective evaluations, further enhancing efficiency. The optimal models identified by the proposed algorithm were evaluated on the CIFAR-10 and CIFAR-100 datasets. The experimental results demonstrate the effectiveness of the proposed algorithm, showing significant improvements in accuracy compared to the existing state-of-the-art methods. This work highlights the potential of evolutionary membrane algorithms to streamline the design and optimization of DSNN architectures, offering a novel and efficient approach to address the challenges in the applications of automated parameter optimization for DSNN.