To check whether or not this drawback holds for in the present day’s massive multimodal fashions, the workforce performed a managed analysis. They skilled the chosen fashions on 5 goal duties, together with fine-grained hen classification, counting, medical visible query answering, OCR studying, and time studying. They then measured how a lot efficiency dropped throughout eight normal benchmarks that weren’t a part of the fine-tuning set.
These experiments led to 2 key discoveries, in response to the paper. Tuning solely the self-attention projection layers (SA Proj), the a part of the mannequin that helps it resolve which enter parts to give attention to, allowed the fashions to study new duties with little or no measurable forgetting. Additionally, what initially appeared as forgotten information typically resurfaced when the mannequin was later skilled on one other specialised activity.
“We thus hypothesize that maybe what seems like forgetting or interference after fine-tuning on a slender goal activity is definitely bias within the output distribution because of the activity distribution shift,” the researchers added. “By way of in-depth evaluation when tuning the counting activity, we verify this speculation: tuning the MLP will increase goal accuracy but additionally will increase the chance of outputting numeric tokens and a extremely correlated drop in held-out activity accuracy, whereas tuning the self-attention achieves the goal studying with out a lot bias towards numeric tokens and with out shedding held-out accuracy.”
