Effective Pre-Training Method and Its Compositional Intelligence for Image Captioning

Abstract

With the increase in the performance of deep learning models, the model parameter has increased exponentially. An increase in model parameters leads to an increase in computation and training time, i.e., an increase in training cost. To reduce the training cost, we propose Compositional Intelligence (CI). This is a reuse method that combines pre-trained models for different tasks. Since the CI uses a well-trained model, good performance and small training cost can be expected in the target task. We applied the CI to the Image Captioning task. Compared to using a trained feature extractor, the caption generator is usually trained from scratch. On the other hand, we pre-trained the Transformer model as a caption generator and applied CI, i.e., we used a pre-trained feature extractor and a pre-trained caption generator. To compare the training cost of the From Scratch model and the CI model, early stopping was applied during fine-tuning of the image captioning task. On the MS-COCO dataset, the vanilla image captioning model reduced training cost by 13.8% and improved performance by up to 3.2%, and the Object Relation Transformer model reduced training cost by 21.3%.