SP-GAN: Self-Growing and Pruning Generative Adversarial Networks.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Traditional Generative Adversarial Networks (GANs) often require large, fixed architectures, leading to training inefficiencies.
• Dynamic network adjustment during training is crucial for optimizing GAN performance and resource utilization.

Purpose of the Study:

• To introduce a novel SP-GAN model capable of dynamically adjusting its architecture during training for realistic image generation.
• To enhance the stability and efficiency of GAN training through self-growing and pruning mechanisms.

Main Methods:

• The SP-GAN utilizes seed networks with minimal convolution kernels, which are then augmented via kernel replication (self-growing).
• A pruning strategy is employed to mitigate excessive network growth and reduce redundancy, optimizing network scale.
• An adaptive loss function with dynamically adjustable hyperparameters is incorporated for improved training dynamics.

Main Results:

• Experimental results demonstrate significant improvements in the stability and efficiency of network training.
• The proposed SP-GAN achieves effective realistic image generation with optimized network architectures.
• The method proves advantageous across various datasets, highlighting its generalizability.

Conclusions:

• The SP-GAN offers a more efficient and stable approach to realistic image generation compared to traditional GAN models.
• Dynamic network architecture adjustment through self-growing and pruning is a viable strategy for enhancing GAN training.
• The adaptive loss function further contributes to the robustness and adaptability of the proposed model.