Spiking generative adversarial network with attention scoring decoding.

Area of Science:

• Deep Learning
• Computational Neuroscience
• Artificial Intelligence

Background:

• Generative models are crucial in deep learning but largely confined to artificial neural networks.
• Spiking neural networks (SNNs), mimicking brain processing, offer potential but have underexplored generative capabilities.
• Existing spiking generative adversarial networks (SGANs) struggle with complex data and performance.

Purpose of the Study:

• To pioneer a high-performance SGAN for complex image generation.
• To address out-of-domain and temporal inconsistencies in SGANs.
• To evaluate the model's performance on diverse static and event-based datasets.

Main Methods:

• Developed a novel spiking generative adversarial network architecture.
• Incorporated Earth-Mover distance to resolve out-of-domain issues.
• Utilized an attention-based weighted decoding method for temporal consistency.

Main Results:

• Achieved state-of-the-art performance on MNIST, FashionMNIST, CIFAR10, and CelebA datasets.
• Demonstrated successful application to event-based data, yielding notable results.
• Outperformed hybrid SGANs and showed closer alignment to mouse brain processing patterns.

Conclusions:

• The proposed SGAN effectively handles complex image generation and overcomes prior limitations.
• This work advances generative modeling in SNNs, showing promise for brain-inspired AI.
• The model offers a more biologically plausible approach to generative tasks.