[论文解读] Supervised Multimodal Bitransformers for Classifying Images and Text

Douwe Kiela, Suvrat Bhooshan+3 more|arXiv (Cornell University)|Sep 6, 2019

Multimodal Machine Learning Applications参考文献 50被引用 163

一句话总结

Introduce a supervised multimodal bitransformer (MMBT) that maps image embeddings into BERT's token space to fuse text and image modalities, achieving competitive results with ViLBERT on text-heavy multimodal classification tasks without multimodal pretraining.

ABSTRACT

Self-supervised bidirectional transformer models such as BERT have led to dramatic improvements in a wide variety of textual classification tasks. The modern digital world is increasingly multimodal, however, and textual information is often accompanied by other modalities such as images. We introduce a supervised multimodal bitransformer model that fuses information from text and image encoders, and obtain state-of-the-art performance on various multimodal classification benchmark tasks, outperforming strong baselines, including on hard test sets specifically designed to measure multimodal performance.

研究动机与目标

• Motivate the need for effective multimodal fusion where text is the dominant modality.
• Propose a simple baseline that finetunes unimodally pretrained text and image encoders for multimodal tasks.
• Show that self-attention over both modalities yields strong performance on text-heavy multimodal classification tasks.
• Demonstrate that the approach is competitive with multimodally pretrained models like ViLBERT while being simpler and extensible.

提出的方法

• Use a ResNet-152 image encoder to produce N image embeddings from KM grid cells.
• Project each image embedding into D-dimensional BERT input space via learned matrices W_n.
• Combine text and image embeddings as contextual embeddings fed into a BERT-like bidirectional transformer initialized from pretrained BERT weights.
• Fine-tune the architecture end-to-end with task-appropriate loss (multiclass cross-entropy or binary cross-entropy for multilabel).
• Handle variable modality presence by segment embeddings and a flexible input layer compatible with multiple modalities.

实验结果

研究问题

• RQ1 Can unimodally pretrained text and image encoders fused through self-attention outperform traditional multimodal fusion baselines on text-heavy multimodal tasks?
• RQ2 How close can a supervised multimodal fusion model approach the performance of self-supervised multimodal pretraining schemes like ViLBERT?
• RQ3 Does freezing/unfreezing of components during fine-tuning affect multimodal fusion quality?
• RQ4 Are the proposed image-to-BERT space mappings robust to missing modalities during inference?
• RQ5 What are the comparative benefits of MMBT versus concatenation-based or gate-based fusion methods on hard multimodal cases?

主要发现

• MMBT outperforms several strong fusion baselines on text-heavy multimodal tasks (MM-IMDB, FOOD101, V-SNLI).
• MMBT is competitive with ViLBERT, sometimes matching or exceeding its performance without multimodal pretraining.
• Hard-subset evaluations show MMBT maintaining strong multimodal performance when unimodal signals conflict.
• Freezing/unfreezing experiments indicate a strategy where the image encoder is unfrozen earlier yields better multimodal integration.
• Constrained-parameter comparisons suggest MMBT can outperform deeper ConcatBert configurations, indicating effective fusion beyond parameter count alone.