What Makes ImageNet Look Unlike LAION

TL;DR

This paper investigates how recreating ImageNet from LAION dataset captions results in a dataset, LAIONet, that differs significantly from the original, highlighting the impact of data collection methods on dataset characteristics and model performance.

Contribution

It introduces LAIONet, a new ImageNet-like dataset created from caption-based searches, and provides a causal explanation for differences in dataset quality and model transferability.

Findings

LAIONet has lower intra-class similarity than ImageNet.

Models trained on ImageNet perform worse on LAIONet.

Caption-based search creates an information bottleneck affecting dataset bias.

Abstract

ImageNet was famously created from Flickr image search results. What if we recreated ImageNet instead by searching the massive LAION dataset based on image captions alone? In this work, we carry out this counterfactual investigation. We find that the resulting ImageNet recreation, which we call LAIONet, looks distinctly unlike the original. Specifically, the intra-class similarity of images in the original ImageNet is dramatically higher than it is for LAIONet. Consequently, models trained on ImageNet perform significantly worse on LAIONet. We propose a rigorous explanation for the discrepancy in terms of a subtle, yet important, difference in two plausible causal data-generating processes for the respective datasets, that we support with systematic experimentation. In a nutshell, searching based on an image caption alone creates an information bottleneck that mitigates the selection bias otherwise present in image-based filtering. Our explanation formalizes a long-held intuition in the community that ImageNet images are stereotypical, unnatural, and overly simple representations of the class category. At the same time, it provides a simple and actionable takeaway for future dataset creation efforts.