UniEvo-VL: Self-Distillation Training for Multimodal Model Self-Improvement

Researchers have introduced UniEvo-VL, a self-distillation training method for multimodal models, which enables self-improvement of these models without relying on human annotations or additional data. UniEvo-VL uses a two-stage training process, where the first stage involves training a teacher model on a large dataset, and the second stage involves training a student model on the output of the teacher model. The student model is then used to refine the teacher model, creating a feedback loop. In experiments, UniEvo-VL demonstrated improved performance on various multimodal tasks, including image classification, object detection, and visual question answering. The method is applicable to various types of multimodal models, including those based on Vision-and-Language BERT (VL-BERT) and VL-BERT with a task-specific head.