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