Probing Conceptual Understanding of Large Visual-Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Schiappa, Madeline, Abdullah, Raiyaan, Azad, Shehreen, Claypoole, Jared, Cogswell, Michael, Divakaran, Ajay, Rawat, Yogesh
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910424918654976
author Schiappa, Madeline
Abdullah, Raiyaan
Azad, Shehreen
Claypoole, Jared
Cogswell, Michael
Divakaran, Ajay
Rawat, Yogesh
author_facet Schiappa, Madeline
Abdullah, Raiyaan
Azad, Shehreen
Claypoole, Jared
Cogswell, Michael
Divakaran, Ajay
Rawat, Yogesh
contents In recent years large visual-language (V+L) models have achieved great success in various downstream tasks. However, it is not well studied whether these models have a conceptual grasp of the visual content. In this work we focus on conceptual understanding of these large V+L models. To facilitate this study, we propose novel benchmarking datasets for probing three different aspects of content understanding, 1) \textit{relations}, 2) \textit{composition}, and 3) \textit{context}. Our probes are grounded in cognitive science and help determine if a V+L model can, for example, determine if snow garnished with a man is implausible, or if it can identify beach furniture by knowing it is located on a beach. We experimented with many recent state-of-the-art V+L models and observe that these models mostly \textit{fail to demonstrate} a conceptual understanding. This study reveals several interesting insights such as that \textit{cross-attention} helps learning conceptual understanding, and that CNNs are better with \textit{texture and patterns}, while Transformers are better at \textit{color and shape}. We further utilize some of these insights and investigate a \textit{simple finetuning technique} that rewards the three conceptual understanding measures with promising initial results. The proposed benchmarks will drive the community to delve deeper into conceptual understanding and foster advancements in the capabilities of large V+L models. The code and dataset is available at: \url{https://tinyurl.com/vlm-robustness}
format Preprint
id arxiv_https___arxiv_org_abs_2304_03659
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Probing Conceptual Understanding of Large Visual-Language Models
Schiappa, Madeline
Abdullah, Raiyaan
Azad, Shehreen
Claypoole, Jared
Cogswell, Michael
Divakaran, Ajay
Rawat, Yogesh
Computer Vision and Pattern Recognition
In recent years large visual-language (V+L) models have achieved great success in various downstream tasks. However, it is not well studied whether these models have a conceptual grasp of the visual content. In this work we focus on conceptual understanding of these large V+L models. To facilitate this study, we propose novel benchmarking datasets for probing three different aspects of content understanding, 1) \textit{relations}, 2) \textit{composition}, and 3) \textit{context}. Our probes are grounded in cognitive science and help determine if a V+L model can, for example, determine if snow garnished with a man is implausible, or if it can identify beach furniture by knowing it is located on a beach. We experimented with many recent state-of-the-art V+L models and observe that these models mostly \textit{fail to demonstrate} a conceptual understanding. This study reveals several interesting insights such as that \textit{cross-attention} helps learning conceptual understanding, and that CNNs are better with \textit{texture and patterns}, while Transformers are better at \textit{color and shape}. We further utilize some of these insights and investigate a \textit{simple finetuning technique} that rewards the three conceptual understanding measures with promising initial results. The proposed benchmarks will drive the community to delve deeper into conceptual understanding and foster advancements in the capabilities of large V+L models. The code and dataset is available at: \url{https://tinyurl.com/vlm-robustness}
title Probing Conceptual Understanding of Large Visual-Language Models
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2304.03659