Actor-Critic Sequence Training for Image Captioning
This work addresses the challenge of improving caption quality for AI agents that need to communicate with humans, though it is incremental as it builds on existing reinforcement learning methods for captioning.
The paper tackles the problem of training image captioning models to directly optimize non-differentiable language quality metrics like CIDEr, rather than just maximizing the likelihood of ground-truth captions. It achieves state-of-the-art performance on the MSCOCO benchmark by using an actor-critic reinforcement learning approach with a per-token advantage and value computation strategy.
Generating natural language descriptions of images is an important capability for a robot or other visual-intelligence driven AI agent that may need to communicate with human users about what it is seeing. Such image captioning methods are typically trained by maximising the likelihood of ground-truth annotated caption given the image. While simple and easy to implement, this approach does not directly maximise the language quality metrics we care about such as CIDEr. In this paper we investigate training image captioning methods based on actor-critic reinforcement learning in order to directly optimise non-differentiable quality metrics of interest. By formulating a per-token advantage and value computation strategy in this novel reinforcement learning based captioning model, we show that it is possible to achieve the state of the art performance on the widely used MSCOCO benchmark.