Multi-Level Collaboration in Model Merging
This work addresses a foundational question in multi-task learning for AI researchers, showing that merging can match ensembling without prior restrictions, though it is incremental in extending theoretical understanding.
The paper tackles the problem of whether model merging can achieve performance consistency with model ensembling under relaxed conditions, and finds that it can attain near-identical and superior performance, as demonstrated by a 95.44% merging accuracy vs. 95.46% ensembling accuracy in experiments with 5 CLIP-ViT-B/32 models.
Parameter-level model merging is an emerging paradigm in multi-task learning with significant promise. Previous research has explored its connections with prediction-level model ensembling-commonly viewed as the upper bound for merging-to reveal the potential of achieving performance consistency between the two. However, this observation relies on certain preconditions, such as being limited to two models, using ViT-based models, and all models are fine-tuned from the same pre-trained checkpoint. To further understand the intrinsic connections between model merging and model ensembling, this paper explores an interesting possibility: If these restrictions are removed, can performance consistency still be achieved between merging and ensembling? To answer this question, we first theoretically establish a performance correlation between merging and ensembling. We find that even when previous restrictions are not met, there is still a way for model merging to attain a near-identical and superior performance similar to that of ensembling. To verify whether our findings are practical, we introduce a validation framework termed Neural Ligand (NeuLig). The learning process of NeuLig is meticulously designed with a specialized loss function supported by theoretical foundations. Experimental results demonstrate the robust resilience of NeuLig in terms of both model scale and the number of collaborating models. For instance, for the case involving 5 CLIP-ViT-B/32 models, parameter-level merging achieves the same performance as prediction-level ensembling (merging: 95.44% vs. ensembling: 95.46%).