Editing Discrete Latent Variables with Reinforcement Learning
An RL agent that makes minimal, interpretable edits to discrete VQ-VAE codes to reach a target class.
This project develops a reinforcement-learning agent that transforms an input toward a specified target class by editing discrete VQ-VAE latent codes. The objective is to reach the desired semantic change while changing as few codebook entries as possible.
The training reward combines classifier confidence with discriminator realism, so the agent learns goal-directed transformations that stay visually plausible. Multi-step editing trajectories show directly how the agent navigates the latent representation, and they reveal task-dependent structure in the codebook.
The broader goal is generative control that is precise and interpretable: a short sequence of discrete decisions instead of an opaque, unconstrained latent update.