Learning a Deep Reinforcement Learning Policy Over the Latent Space of a Pre-trained GAN for Semantic Age Manipulation

Learning a disentangled representation of the latent space has become one of\nthe most fundamental problems studied in computer vision. Recently, many\nGenerative Adversarial Networks (GANs) have shown promising results in\ngenerating high fidelity images. However, studies to understand the semantic\nlayout of the latent space of pre-trained models are still limited. Several\nworks train conditional GANs to generate faces with required semantic\nattributes. Unfortunately, in these attempts, the generated output is often not\nas photo-realistic as the unconditional state-of-the-art models. Besides, they\nalso require large computational resources and specific datasets to generate\nhigh fidelity images. In our work, we have formulated a Markov Decision Process\n(MDP) over the latent space of a pre-trained GAN model to learn a conditional\npolicy for semantic manipulation along specific attributes under defined\nidentity bounds. Further, we have defined a semantic age manipulation scheme\nusing a locally linear approximation over the latent space. Results show that\nour learned policy samples high fidelity images with required age alterations,\nwhile preserving the identity of the person.\n