MetaReVision

This repository is still actively under construction!

This is the official implementation of:

MetaReVision: Meta-Learning with Retrieval for Visually Grounded Compositional Concept Acquisition
Guangyue Xu, Parisa Kordjamshidi, Joyce Chai Michigan State University, University of Michigan EMNLP-Finding 2023

Project Page | Paper

Train MetaReVision

Step 1 - Pretrain vision-language models (VLMs)

Step 2 - Using the pretrained VLMs to construct element-concept DB

Step 3 - Construct episodes for Compositional Learning using the above DB

Step 4 - Meta-tune VLMs using the constructed Episodes

Citation

@inproceedings{xu2023metarevision,
    title={MetaReVision: Meta-Learning with Retrieval for Visually Grounded Compositional Concept Acquisition}, 
    author={Guangyue Xu and Parisa Kordjamshidi and Joyce Chai},
    year={2023},
    booktitle={Findings of the Association for Computational Linguistics: EMNLP 2023}
}