Data, Architecture, or Losses: What Contributes Most to Multimodal Transformer Success?

Recently multimodal transformer models have gained popularity because their performance on language and vision tasks suggests they learn rich visual-linguistic representations. Focusing on zero-shot image retrieval tasks, we study three important factors which can impact the quality of learned representations: pretraining data, the attention mechanism, and loss functions. By pretraining models on six datasets, we observe that dataset noise and language similarity to our downstream task are important indicators of model performance. Through architecture analysis, we learn that models with a multimodal attention mechanism can outperform deeper models with modality-specific attention mechanisms. Finally, we show that successful contrastive losses in self-supervised learning work do not lead to similar performance gains when used on multimodal transformers.