Decoupling the Role of Data, Attention, and Losses in Multimodal Transformers
This work addresses the problem of optimizing multimodal transformers for researchers and practitioners, but it is incremental as it analyzes existing factors rather than introducing new methods.
The study investigated how pretraining data, attention mechanisms, and loss functions affect multimodal transformer performance in zero-shot image retrieval, finding that dataset noise and language similarity are key indicators, multimodal attention outperforms deeper modality-specific models, and contrastive losses do not improve performance.
Recently multimodal transformer models have gained popularity because their performance on language and vision tasks suggest 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 architectural 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 used in the self-supervised learning literature do not yield similar performance gains when used in multimodal transformers