Impact of Pretraining Word Co-occurrence on Compositional Generalization in Multimodal Models

TL;DR

This paper investigates how word co-occurrence statistics in training data influence the compositional generalization of multimodal models like CLIP, revealing strong correlations between PMI and zero-shot accuracy across synthetic and natural images.

Contribution

It demonstrates that word co-occurrence, measured by PMI, significantly impacts model accuracy and transfers across different multimodal models, highlighting the need for improved compositional generalization methods.

Findings

Strong correlation (r=0.97) between PMI and CLIP accuracy on synthetic images.

Reproduced PMI-accuracy correlation (r=0.75) in natural images through editing.

Transfer of PMI-accuracy relationship observed in LMMs built on CLIP.

Abstract

CLIP and large multimodal models (LMMs) have better accuracy on examples involving concepts that are highly represented in the training data. However, the role of concept combinations in the training data on compositional generalization is largely unclear -- for instance, how does accuracy vary when a common object appears in an uncommon pairing with another object? In this paper, we investigate how word co-occurrence statistics in the pretraining dataset (a proxy for co-occurrence of visual concepts) impacts CLIP/LMM performance. To disentangle the effects of word co-occurrence frequencies from single-word frequencies, we measure co-occurrence with pointwise mutual information (PMI), which normalizes the joint probability of two words co-occurring by the probability of co-occurring independently. Using synthetically generated images with a variety of concept pairs, we show a strong correlation between PMI in the CLIP pretraining data and zero-shot accuracy in CLIP models trained on LAION-400M (r=0.97 and 14% accuracy gap between images in the top and bottom 5% of PMI values), demonstrating that even accuracy on common concepts is affected by the combination of concepts in the image. Leveraging this finding, we reproduce this effect in natural images by editing them to contain pairs with varying PMI, resulting in a correlation of r=0.75. Finally, we demonstrate that this behavior in CLIP transfers to LMMs built on top of CLIP (r=0.70 for TextVQA, r=0.62 for VQAv2). Our findings highlight the need for algorithms and architectures that improve compositional generalization in multimodal models without scaling the training data combinatorially. Our code is available at https://github.com/helenqu/multimodal-pretraining-pmi.