Detrimental Memories in Transfer Learning

Amal Alnouri, Timothy J Wroge, Bilal Alsallakh

Keywords: Transfer Learning, Negative Transfer, Finetuning, Foundational Models

TL;DR: We propose lingering memories about the source domain as the mechanism behind negative transfer learning that explains odd failure cases of Foundational Models.

Abstract: The source domain in transfer learning provides essential features that enable effective and data-efficient learning on the target task. Typically, the finetuning process does not explicitly account for how the knowledge about the source domain interacts with the target task. We demonstrate how that knowledge can interfere with the target task leading to negative transfer. Specifically, certain memories about the source domain can distract the finetuned model in certain inputs. We provide a method to analyze those memories in typical foundational models and to surface potential failure cases of those models. This analysis helps model developers explore remedies for those failure cases, such as expanding the training data or adapting the training objective.

Submission Number: 66