Adaptive Kernels in DCGANs

Abstract

Deep Convolutional Generative Adversarial Networks (DCGANs) is a specialized deep learning architecture tailored for image generation tasks. DCGANs have notably enhanced the field of image generation by introducing robust training techniques and architectural principles that yield high-fidelity images that closely mirror those present in the training dataset. Recent research suggests that employing a deterministic adaptive kernel postconvolution can bolster CNN’s generalization capabilities, thereby benefiting DCGANs. However, the challenge of generating the weights for these deterministic adaptive kernels remains an active area of research. In response, we propose an innovative adaptive kernel method that utilizes the convoluted data from a layer to generate a dynamic set of four Gaussian kernels. Subsequently, convolution operations are performed on the convoluted data using either a single Gaussian kernel at a time or all four kernels sequentially, with the order of application rotating after each epoch. To validate our novel adaptive kernel approach, we conducted experiments using DCGAN and one of its variants, evaluating performance on two diverse datasets (CIFAR-10 and CIFAR-100) for image generation tasks. Importantly, our method seamlessly integrates with any DCGAN variant as a plug-and-play solution, introducing no additional trainable parameters to the network. We also offer a comprehensive analysis of the impact of our proposed adaptive kernel method.