Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MM-SHAP: A Performance-agnostic Metric for Measuring Multimodal Contributions in Vision and Language Models & Tasks

Letitia Parcalabescu, Anette Frank

TL;DR

A performance-agnostic multimodality score based on Shapley values that reliably quantifies in which proportions a multimodal model uses individual modalities is proposed, contradicting the wide-spread assumption that unimodal collapse is one-sided.

Abstract

Vision and language models (VL) are known to exploit unrobust indicators in individual modalities (e.g., introduced by distributional biases) instead of focusing on relevant information in each modality. That a unimodal model achieves similar accuracy on a VL task to a multimodal one, indicates that so-called unimodal collapse occurred. However, accuracy-based tests fail to detect e.g., when the model prediction is wrong, while the model used relevant information from a modality. Instead, we propose MM-SHAP, a performance-agnostic multimodality score based on Shapley values that reliably quantifies in which proportions a multimodal model uses individual modalities. We apply MM-SHAP in two ways: (1) to compare models for their average degree of multimodality, and (2) to measure for individual models the contribution of individual modalities for different tasks and datasets. Experiments with six VL models -- LXMERT, CLIP and four ALBEF variants -- on four VL tasks highlight that unimodal collapse can occur to different degrees and in different directions, contradicting the wide-spread assumption that unimodal collapse is one-sided. Based on our results, we recommend MM-SHAP for analysing multimodal tasks, to diagnose and guide progress towards multimodal integration. Code available at \url{https://github.com/Heidelberg-NLP/MM-SHAP}.

MM-SHAP: A Performance-agnostic Metric for Measuring Multimodal Contributions in Vision and Language Models & Tasks

TL;DR

A performance-agnostic multimodality score based on Shapley values that reliably quantifies in which proportions a multimodal model uses individual modalities is proposed, contradicting the wide-spread assumption that unimodal collapse is one-sided.

Abstract

Vision and language models (VL) are known to exploit unrobust indicators in individual modalities (e.g., introduced by distributional biases) instead of focusing on relevant information in each modality. That a unimodal model achieves similar accuracy on a VL task to a multimodal one, indicates that so-called unimodal collapse occurred. However, accuracy-based tests fail to detect e.g., when the model prediction is wrong, while the model used relevant information from a modality. Instead, we propose MM-SHAP, a performance-agnostic multimodality score based on Shapley values that reliably quantifies in which proportions a multimodal model uses individual modalities. We apply MM-SHAP in two ways: (1) to compare models for their average degree of multimodality, and (2) to measure for individual models the contribution of individual modalities for different tasks and datasets. Experiments with six VL models -- LXMERT, CLIP and four ALBEF variants -- on four VL tasks highlight that unimodal collapse can occur to different degrees and in different directions, contradicting the wide-spread assumption that unimodal collapse is one-sided. Based on our results, we recommend MM-SHAP for analysing multimodal tasks, to diagnose and guide progress towards multimodal integration. Code available at \url{https://github.com/Heidelberg-NLP/MM-SHAP}.
Paper Structure (48 sections, 3 equations, 11 figures, 3 tables)

This paper contains 48 sections, 3 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Testing for unimodal collapse
  4. Interpretability
  5. Quantifying Multimodal Contributions
  6. A case for a performance-agnostic score
  7. Background on Shapley Values
  8. MM-SHAP
  9. Why SHAP enables a MM score
  10. Ways of using MM-SHAP
  11. Sample-level
  12. Dataset and model level
  13. Measuring fine-tuning effects
  14. Multimodal Contributions across Models and Datasets
  15. Tasks
  16. ...and 33 more sections

Figures (11)

  • Figure 1: We display image-sentence alignment scores (ISA) and the textual degreeT-SHAP that measures how much models focus on text rather than the image (with $100 - \texttt{T-SHAP}$% the corresponding visual degree) for 3 VL models. Blue/red highlights on text tokens and image tokens (patches) contribute towards higher/lower ISA. Note: CLIP's ISA is an absolute score, while ALBEF and LXMERT predict ISA probabilities. See Section \ref{['sec:exps-results']} for more details on this figure; App. \ref{['app:samples']} for more detailed analysis of this instance and more samples.
  • Figure 2: Low discrepancynoun phrase foil: Image-sentence alignment score (ISA) of the six VL models with their textual degree T-SHAP (in %). Each text and image token (image patch) is colour-coded: Blue tokens contribute to a high ISA, while red ones lower the ISA. The visual degree is $100 - \texttt{T-SHAP}$. Note that the ISA of CLIP is an absolute score, while ALBEF and LXMERT predict ISA probabilities. With *◯ we mark correct ISA and highlight the correct / foil token that contributes in the right direction for aligning the image and the caption. With *◯, we mark incorrect ISA and wrong contribution directions.
  • Figure 3: Low discrepancy (VALSE*◯action replacement): Image-sentence alignment score (ISA) of the six VL models with their textual degree T-SHAP (in %). Each text and image token (image patch) is colour-coded: Blue tokens contribute to a high ISA, while red ones lower the ISA. The visual degree is $100 - \texttt{T-SHAP}$. Note that the ISA of CLIP is an absolute score, while ALBEF and LXMERT predict ISA probabilities. With *◯ we mark correct ISA and an highlight the correct / foil token that contributes in the right direction for aligning the image and the caption. With *◯, we mark incorrect ISA and wrong contribution directions.
  • Figure 4: Low discrepancy (VALSE*◯counting): Image-sentence alignment score (ISA) of the six VL models with their textual degree T-SHAP (in %). Each text and image token (image patch) is colour-coded: Blue tokens contribute to a high ISA, while red ones lower the ISA. The visual degree is $100 - \texttt{T-SHAP}$. Note that the ISA of CLIP is an absolute score, while ALBEF and LXMERT predict ISA probabilities. With *◯ we mark correct ISA and an highlight the correct / foil token that contributes in the right direction for aligning the image and the caption. With *◯, we mark incorrect ISA and wrong contribution directions.
  • Figure 5: Low discrepancy (VALSE*◯existence positive): Image-sentence alignment score (ISA) of the six VL models with their textual degree T-SHAP (in %). Each text and image token (image patch) is colour-coded: Blue tokens contribute to a high ISA, while red ones lower the ISA. The visual degree is $100 - \texttt{T-SHAP}$. Note that the ISA of CLIP is an absolute score, while ALBEF and LXMERT predict ISA probabilities. With *◯ we mark correct ISA and an highlight the correct / foil token that contributes in the right direction for aligning the image and the caption. With *◯, we mark incorrect ISA and wrong contribution directions.
  • ...and 6 more figures