科研成果详情

Spiking neural networks (SNNs) have gained popularity due to their ability to operate at ultra-high speed and ultra-low power consumption, making them suitable for energy-efficient applications in various fields. However, current SNNs lack the capability to generate high-quality images, and SNN-based generative models often confront the issue of catastrophic forgetting when dealing with sequential probabilistic representations of multiple datasets. In this work, we propose a novel SNN-based generative framework called Lifelong-SGN, which consists of three components: Teacher, Student, and Assistant. To overcome the issue of catastrophic forgetting, we utilize the Teacher module based on Generative Adversarial Network (GAN), to store and replay probabilistic representations from previously learned knowledge. For the Student module, we utilize an SNN-based VAE implementation to learn probabilistic representations from both the output obtained by the Teacher module and the new dataset. Moreover, we introduce the Discriminator module as an Assistant to train the SNN-based generative model using adversarial training principles, which helps generate high-quality images. Our experiments on Lifelong-SGN demonstrate its effectiveness in image classification and generation in lifelong learning environment.