Data Descriptions from Large Language Models with Influence Estimation
Keywords: Explainable AI, multi-modal, vision-language model, CLIP, cross-modal transferability, image classification, large language models, 2-stage prompting
Abstract: Deep learning models have been successful in many areas but understanding their behaviors still remains a black-box. Most prior explainable AI (XAI) approaches have focused on interpreting and explaining how models make predictions. But in contrast, we take a different approach via the lens of data because data is one of the most important factors in the success of deep learning models. We would like to understand how data can be explained with deep learning model training via one of the most common media -- language. Therefore, we propose a novel approach to understand and extract which information can explain each class inside the dataset well by incorporating knowledge from existing external knowledge bases extracted through large language models such as GPT-3.5. However, the extracted data descriptions may still include irrelevant information, so we propose to exploit influence estimation to choose the most informative textual descriptions, along with the CLIP score. The presented textual descriptions may provide insight into what the trained model focuses on and utilizes for making the prediction. Furthermore, by utilizing recent vision-language contrastive learning as it may provide cross-modal transferability, we propose a novel benchmark task of cross-modal transfer classification to examine the effectiveness of the data description. In experiments with nine image classification datasets, the extracted text descriptions further boost the performance of the trained model with only images. Therefore, it demonstrates that the proposed approach provides information that can explain the characteristics of each dataset that helps the model to train. Through this, we may have insight and inherent interpretability of the decision process from the model. In addition, we show that our approach solves model bias in text-to-image generation tasks.
Primary Area: visualization or interpretation of learned representations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5006
Loading