Decoupling the Role of Data, Attention, and Losses in Multimodal Transformers
arXiv:2102.00529
Abstract
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
arXiv commentpre-print of MIT Press Publication version
References in corpus (11)
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- Language Models are Few-Shot Learners
- Microsoft COCO Captions: Data Collection and Evaluation Server
- Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models
- VisualBERT: A Simple and Performant Baseline for Vision and Language
- Hierarchical Question-Image Co-Attention for Visual Question Answering
- HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips
- Contrastive Multiview Coding
- Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers
- Learning a Text-Video Embedding from Incomplete and Heterogeneous Data
- Are we pretraining it right? Digging deeper into visio-linguistic pretraining