Student Can Also be a Good Teacher: Extracting Knowledge from Vision-and-Language Model for Cross-Modal Retrieval

Astounding results from transformer models with Vision-and Language Pretraining (VLP) on joint vision-and-language downstream tasks have intrigued the multi-modal community. On the one hand, these models are usually so huge that make us more difficult to fine-tune and serve real-time online applications. On the other hand, the compression of the original transformer block will ignore the difference in information between modalities, which leads to the sharp decline of retrieval accuracy. In this work, we present a very light and effective cross-modal retrieval model compression method. With this method, by adopting a novel random replacement strategy and knowledge distillation, our module can learn the knowledge of the teacher with nearly the half number of parameters reduction. Furthermore, our compression method achieves nearly 130x acceleration with acceptable accuracy. To overcome the sharp decline in retrieval tasks because of compression, we introduce the co-attention interaction module to reflect the different information and interaction information. Experiments show that a multi-modal co-attention block is more suitable for cross-modal retrieval tasks rather than the source transformer encoder block.