# Meta Learning to Bridge Vision and Language Models for Multimodal Few-Shot Learning

**Authors:** Ivona Najdenkoska, Xiantong Zhen, Marcel Worring  
**arXiv:** [2302.14794](https://arxiv.org/abs/2302.14794)  
**Date:** 2023-02-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.14794v1/

## Key points

- Paper title: "Meta Learning to Bridge Vision and Language Models for Multimodal Few-Shot Learning"
- Addresses the challenge of multimodal few-shot learning
- Proposes a novel meta-learning approach
- Bridges the domain gap between vision and language modalities
- Existing methods rely on hand-engineered task induction and prompts to frozen language models, limiting performance
- Proposed method decomposes model training into related multimodal few-shot tasks
- Introduces a meta-mapper network as a meta-learner
- Meta-mapper acquires shared meta-knowledge across tasks by updating learnable parameters only
- Enables rapid adaptation to new samples with just a few gradient updates
- Induces tasks in a data-driven manner without requiring hand-engineered task induction
- Experimental results demonstrate superior performance and computational efficiency compared to existing approaches
- Presents a promising solution for multimodal few-shot learning by leveraging shared meta-knowledge among related tasks through a novel meta-learning approach.

## Summary

The paper titled "Meta Learning to Bridge Vision and Language Models for Multimodal Few-Shot Learning" addresses the challenge of multimodal few-shot learning by proposing a novel meta-learning approach. This approach effectively bridges the significant domain gap between vision and language modalities, which has hindered progress in this field. Existing methods rely on hand-engineered task induction and prompts to frozen language models, limiting their performance. In contrast, the proposed method decomposes model training into related multimodal few-shot tasks and introduces a meta-mapper network as a meta-learner. By updating only learnable parameters, the meta-mapper acquires shared meta-knowledge across tasks, enabling rapid adaptation to new samples with just a few gradient updates. Importantly, this approach induces tasks in a data-driven manner without requiring hand-engineered task induction. Experimental results on various benchmarks demonstrate the superior performance and computational efficiency of the proposed method compared to existing approaches. In summary, this paper presents a promising solution for multimodal few-shot learning by leveraging shared meta-knowledge among related tasks through a novel meta-learning approach.