Web-crawled pretraining datasets underlie the impressive "zero-shot" evaluation performance of multimodal models, such as CLIP for classification/retrieval and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of "zero-shot" generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during "zero-shot" evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets? We comprehensively investigate this question across 34 models and five standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting "zero-shot" generalization, multimodal models require exponentially more data to achieve linear improvements in downstream "zero-shot" performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the "Let it Wag!" benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to "zero-shot" generalization capabilities under large-scale training paradigms remains to be found.
Zero-shot learning (ZSL) is a machine learning scenario in which an AI model is trained to recognize and categorize objects or concepts without having seen any examples of those categories or concepts beforehand.
Most state-of-the-art deep learning models for classification or regression are trained through supervised learning, which requires many labeled examples of relevant data classes. Models “learn” by making predictions on a labeled training dataset; data labels provide both the range of possible answers and the correct answers (or ground truth) for each training example.
While powerful, supervised learning is impractical in some real-world scenarios. Annotating large amounts of data samples is costly and time-consuming, and in cases like rare diseases and newly discovered species, examples may be scarce or non-existent.
so yeah, i agree, the paper is saying these models aren't capable of creating/using human-understandable concepts without gobs and gobs of training data, and if you try to take human supervision of those categories out of the process, then you need even more gobs and gobs of training data. edge cases and novel categories tend to spin off useless bullshit from these things.
but there's some speculation that the recent stock market downturn affecting tech stocks especially may be related to the capitalist class figuring out that these things aren't actually magical knowledge-worker replacement devices and won't let them make the line go up forever and ever amen. so even if the suits don't really digest the contents of this paper, they'll figure out the relevant parts reventually.