Probing Conceptual Understanding of Large Visual-Language Models

TL;DR

This work introduces three cognitive-science-inspired benchmarks—Probe-R, Probe-C, and Probe-B—to probe relational, compositional, and contextual understanding in large visual-language models. Evaluations across ten state-of-the-art V+L models reveal that many models struggle to exhibit robust conceptual understanding, with cross-attention improving relational learning and CNN vs ViT backbones showing complementary strengths in texture vs color/shape. A simple finetuning paradigm using selective negatives (RelComp) demonstrates promise for enhancing compositional and relational reasoning, albeit with some trade-offs in downstream accuracy. The proposed datasets and findings aim to guide the development of V+L systems with grounded conceptual maps, enabling more reliable generalization and transfer across tasks.

Abstract

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}

Figures (13)

• Figure 1: Overview of proposed benchmarks. Probe-R swaps the real subject or relation with an unlikely one and swaps a set of subject-only images to a subject-only prompt and the ground-truth relation prompt. Probe-C asks the model to match two images and two prompts, swapping object or composition. Probe-B compares object recognition performance before and after swapping out context from background and other surrounding objects.
• Figure 2: Model's performance on relational understanding on Probe-R. (left) Radar plot showing accuracy and mean confidence $\mu(c)$ of different models. Here, the anchor image $X_{R_1}$ contains the relation $R_1=\langle s, r, o \rangle$, image $X_{O_1}$ contains $O_1 = \langle s \rangle$. Prompts contain either the relation $P_{R_1}$, $P_{R_2}=\langle s, \overline{r}, o \rangle$, $P_{R_3}=\langle \overline{s}, r, o \rangle$, or $P_{O_1}=\langle s \rangle$. (right) TSNE plot of the feature space for image features for some models where the prompt with the predicate swapped is denoted by $P_{R_2}$ and the ground truth prompt denoted by $P_{R_1}$.
• Figure 3: Model's performance on compositional understanding on Probe-C. (left) The overall results for Probe-C showing the image, text, and group scores for when the object is swapped (Obj.) or when the composition is swapped (Comp.). (middle) Mean group score averaged across attribute categories. (right) CLIP scores averaged over different backbones.
• Figure 4: Model's performance on contextual understanding on Probe-B for only background removal. (left) Mean results for replacing background with filler and (right) for each model averaged over fillers. Comparisons between the original $x_0$, original+random patch $\tilde{x_0}$ and modified $\tilde{x_1}$. The metrics are mAP and $\gamma^r$.
• Figure 5: Model's performance on contextual understanding on Probe-B on background and all but one object removal.(Left): Results for when the background and all other objects are replaced with a filler $\tilde{x_1}$, compared to the original $x_0$, and (right) original+random patch $\tilde{x_0}$.