To check whether or not this drawback holds for at present’s giant multimodal fashions, the staff carried out a managed analysis. They educated the chosen fashions on 5 goal duties, together with fine-grained chicken classification, counting, medical visible query answering, OCR studying, and time studying. They then measured how a lot efficiency dropped throughout eight commonplace benchmarks that weren’t a part of the fine-tuning set.
These experiments led to 2 key discoveries, in keeping with the paper. Tuning solely the self-attention projection layers (SA Proj), the a part of the mannequin that helps it determine which enter components to deal with, allowed the fashions to be taught new duties with little or no measurable forgetting. Additionally, what initially appeared as forgotten data usually resurfaced when the mannequin was later educated on one other specialised activity.
“We thus hypothesize that maybe what appears like forgetting or interference after fine-tuning on a slim goal activity is definitely bias within the output distribution as a result of activity distribution shift,” the researchers added. “Via in-depth evaluation when tuning the counting activity, we verify this speculation: tuning the MLP will increase goal accuracy but in addition 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 dropping held-out accuracy.”
