Base-Change at Prediction: Inference-Time Update of Fine-Tuned Models

Daiki Chijiwa, Taku Hasegawa, Kyosuke Nishida, Kuniko Saito, Susumu Takeuchi

Keywords: Foundation Models, Fine-Tuning

TL;DR: We investigate how to enforce fine-tuned models to keep up with updates of the underlying pre-trained model, without actual training.

Abstract: Foundation models play a central role in recent developments of artificial intelligence on both vision and language domains. However, even if a foundation model is powerful enough at the time to be fine-tuned for various tasks, it will be eventually outdated due to its old knowledge or inadequate capability for new tasks, and then a new foundation model will be prepared by re-training the outdated model with updated data. As a result, the various fine-tuned models based on the outdated model also have to keep up with the new foundation model, typically by fine-tuning again the new foundation model for each task, which should be costly if the number of fine-tuned models or the frequency of updates increases. In this paper, with our simplified theoretical framework, we first derive a probabilistic formula for the fine-tuned model of the new foundation model. Then, based on the formula, we propose a method to avoid the fine-tuning of new foundation models, by editing the predictions of the fine-tuned model in direction to the new foundation model. Compared to previous methods, which edit the predictions of the new foundation model instead, our method consistently keeps or improves accuracy of fine-tuned model for various tasks.

Submission Number: 50