aiqbench analysis
This study reveals a critical trade-off in model deployment where fine-tuning large-scale foundation models often leads to catastrophic forgetting of out-of-distribution robustness, particularly when the pretraining dataset is excessively large. The authors introduce the ImageNet-RIB benchmark to quantify this phenomenon, demonstrating that smaller supervised datasets can sometimes preserve generalization capabilities better than massive contrastive learning corpora like LAION-2B. These findings challenge the assumption that scaling pretraining data linearly improves downstream task resilience and suggest a need for more nuanced strategies in model selection and fine-tuning protocols.
Large Pretraining Datasets Don't Guarantee Robustness after Fine-Tuning in Image Classification
arXiv:2410.21582v4 Announce Type: replace-cross Abstract: Large-scale pretrained models are widely leveraged as foundations for learning new specialized tasks via fine-tuning, with the goal of maintaining the general performance of the model while allowing it to gain new skills. A valuable goal for all such models is robustness: the ability to perform well on out-of-distribution (OOD) tasks. We assess whether fine-tuning preserves the overall robustness of the pretrained model in image classification, and observed that models pretrained on large datasets exhibited strong catastrophic forgetting and loss of OOD generalization. To systematically assess robustness preservation in fine-tuned models, we propose the Robustness Inheritance Benchmark (ImageNet-RIB). The benchmark, which can be applied to any pretrained model, consists of a set of related but distinct OOD (downstream) tasks and involves fine-tuning on one of the OOD tasks in the set then testing on the rest. We find that though continual learning methods help, fine-tuning reduces robustness across pretrained models. Surprisingly, models pretrained on the largest and most diverse datasets (e.g., LAION-2B) exhibit both larger robustness losses and lower absolute robustness after fine-tuning on small datasets, relative to models pretrained on smaller datasets. We observe this collapse in contrastively pretrained (CLIP) models and their fine-tuned variants, where it grows with pretraining scale; the supervised models we test do not exhibit it. These findings suggest that starting with the strongest foundation model is not necessarily the best approach for performance on specialist tasks. https://jd730.github.io/projects/ImageNet-RIB
Abstract:Large-scale pretrained models are widely leveraged as foundations for learning new specialized tasks via fine-tuning, with the goal of maintaining the general performance of the model while allowing it to gain new skills. A valuable goal for all such models is robustness: the ability to perform well on out-of-distribution (OOD) tasks. We assess whether fine-tuning preserves the overall robustness of the pretrained model in image classification, and observed that models pretrained on large datasets exhibited strong catastrophic forgetting and loss of OOD generalization. To systematically assess robustness preservation in fine-tuned models, we propose the Robustness Inheritance Benchmark (ImageNet-RIB). The benchmark, which can be applied to any pretrained model, consists of a set of related but distinct OOD (downstream) tasks and involves fine-tuning on one of the OOD tasks in the set then testing on the rest. We find that though continual learning methods help, fine-tuning reduces robustness across pretrained models. Surprisingly, models pretrained on the largest and most diverse datasets (e.g., LAION-2B) exhibit both larger robustness losses and lower absolute robustness after fine-tuning on small datasets, relative to models pretrained on smaller datasets. We observe this collapse in contrastively pretrained (CLIP) models and their fine-tuned variants, where it grows with pretraining scale; the supervised models we test do not exhibit it. These findings suggest that starting with the strongest foundation model is not necessarily the best approach for performance on specialist tasks. this https URL
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2410.21582 [cs.CV]
(or arXiv:2410.21582v4 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2410.21582 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 333 words
Related coverage
PIRAMID: Progress and Plans
LessWrong · Jul 27, 2026
How might continual learning affect safety and alignment?
LessWrong · Jun 13, 2026
Small LLMs perform no-CoT two-hop reasoning on synthetic facts with real bridges
LessWrong · Oct 17, 2026
Building a Foundation Stack for General-Purpose Robots
IEEE Spectrum · Jul 13, 2026