Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Is the Modality Gap a Bug or a Feature? A Robustness Perspective

Rhea Chowers, Oshri Naparstek, Udi Barzelay, Yair Weiss

Abstract

Many modern multi-modal models (e.g. CLIP) seek an embedding space in which the two modalities are aligned. Somewhat surprisingly, almost all existing models show a strong modality gap: the distribution of images is well-separated from the distribution of texts in the shared embedding space. Despite a series of recent papers on this topic, it is still not clear why this gap exists nor whether closing the gap in post-processing will lead to better performance on downstream tasks. In this paper we show that under certain conditions, minimizing the contrastive loss yields a representation in which the two modalities are separated by a global gap vector that is orthogonal to their embeddings. We also show that under these conditions the modality gap is monotonically related to robustness: decreasing the gap does not change the clean accuracy of the models but makes it less likely that a model will change its output when the embeddings are perturbed. Our experiments show that for many real-world VLMs we can significantly increase robustness by a simple post-processing step that moves one modality towards the mean of the other modality, without any loss of clean accuracy.

Is the Modality Gap a Bug or a Feature? A Robustness Perspective

Abstract

Many modern multi-modal models (e.g. CLIP) seek an embedding space in which the two modalities are aligned. Somewhat surprisingly, almost all existing models show a strong modality gap: the distribution of images is well-separated from the distribution of texts in the shared embedding space. Despite a series of recent papers on this topic, it is still not clear why this gap exists nor whether closing the gap in post-processing will lead to better performance on downstream tasks. In this paper we show that under certain conditions, minimizing the contrastive loss yields a representation in which the two modalities are separated by a global gap vector that is orthogonal to their embeddings. We also show that under these conditions the modality gap is monotonically related to robustness: decreasing the gap does not change the clean accuracy of the models but makes it less likely that a model will change its output when the embeddings are perturbed. Our experiments show that for many real-world VLMs we can significantly increase robustness by a simple post-processing step that moves one modality towards the mean of the other modality, without any loss of clean accuracy.

Paper Structure

This paper contains 33 sections, 5 theorems, 52 equations, 19 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Theory
  4. Notation and Definitions
  5. Why Should a Global Gap Exist?
  6. The Modality Gap Decreases Robustness
  7. Algorithm
  8. Experiments
  9. Robustness Under Controlled Noise
  10. Robustness to Quantization
  11. Robustness to Rephrasing
  12. Discussion
  13. Theorem Proofs
  14. Proof of \ref{['theorem:newtheory']}
  15. Proof of \ref{['theorem:mmloss']}
  16. ...and 18 more sections

Key Result

Theorem 3.1

Let $\mu_x$ and $\mu_y$ be the modalities' empirical means. Assume that $\forall i: \left\lVert y_i-\mu_y\right\rVert \leq \epsilon, \left\lVert x_i-\mu_x\right\rVert \leq \epsilon$, and $\left\lVert\mu_x - \mu_y\right\rVert\gg\epsilon$. Then the gradient of the contrastive loss with respect to the

Figures (19)

  • Figure 1: Left: Projections of CLIPs embedding of the MS-COCO validation set onto its first 3 principal components. A clear separation between images and texts is evident. Right: An image from the Imagenet imagenet validation set that's misclassified by CLIP when changing the caption template. Multi-modal models can lose more than $6\%$ of their accuracy when replacing the caption template.
  • Figure 2: Is the modality gap a bug or a feature? Changing the gap by moving the text embeddings by $\alpha\cdot\vec{g}$ has an inconsistent effect on downstream performance. The figure follows liang2022mind and shows that for some datasets, models benefit from slightly enlarging the gap (\ref{['fig:naive_gapinet']}), some from maintaining it (\ref{['fig:naive_gapcoco']}) .
  • Figure 3: Three points in each modality in $\mathbb{R}^2$ and the corresponding multi-modal contrastive loss along with the magnitude of the gradient of the loss. Lines connect true pairs. As long as the points satisfy relative alignment - the true pair of any point is also its nearest neighbor - the loss and the gradient magnitude are close to zero, even when there exists a gap.
  • Figure 4: The evolution of embeddings using gradient descent on the contrastive loss (\ref{['mmloss']}) starting with two tight clusters. Both when using an unnormalized embedding space (top) and when constricting the embeddings to the sphere (bottom) they converge to a solution that has almost zero loss and for which a global gap vector exists and the gap vector is orthogonal to both modalities. We prove that minimizing the contrastive loss will lead to such a solution under certain assumptions. For full training details see \ref{['appendix:simulations']}.
  • Figure 5: (Top:) Two initial embeddings where the bottom embedding is color-coded based on $S^y_i$. $S^y_i$ decreases with distance to the other modality. ( Bottom:) The training dynamics of a toy model initialized with isotropic Gaussians. Training starts by shrinking variance in the direction of the gap according to \ref{['theorem:newtheory']}.
  • ...and 14 more figures

Theorems & Definitions (10)

  • Theorem 3.1
  • Theorem 3.2
  • Theorem 3.4
  • Theorem 3.5
  • proof
  • proof
  • Lemma A.1
  • proof
  • proof
  • proof