REVIEW 4 cited by
Transformer Fusion with Optimal Transport
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Fusion is a technique for merging multiple independently-trained neural networks in order to combine their capabilities. Past attempts have been restricted to the case of fully-connected, convolutional, and residual networks. This paper presents a systematic approach for fusing two or more transformer-based networks exploiting Optimal Transport to (soft-)align the various architectural components. We flesh out an abstraction for layer alignment, that can generalize to arbitrary architectures - in principle - and we apply this to the key ingredients of Transformers such as multi-head self-attention, layer-normalization, and residual connections, and we discuss how to handle them via various ablation studies. Furthermore, our method allows the fusion of models of different sizes (heterogeneous fusion), providing a new and efficient way to compress Transformers. The proposed approach is evaluated on both image classification tasks via Vision Transformer and natural language modeling tasks using BERT. Our approach consistently outperforms vanilla fusion, and, after a surprisingly short finetuning, also outperforms the individual converged parent models. In our analysis, we uncover intriguing insights about the significant role of soft alignment in the case of Transformers. Our results showcase the potential of fusing multiple Transformers, thus compounding their expertise, in the budding paradigm of model fusion and recombination. Code is available at https://github.com/graldij/transformer-fusion.
Forward citations
Cited by 4 Pith papers
-
Decentralized SGD with Controlled Disagreement Finds Flatter Minima
Keeping consensus errors alive in decentralized SGD via a learning-rate-scaled mixing term improves test accuracy and flatter minima over both DSGD and synchronous SGD.
-
Model Fusion via Retrofitting
A neuron-centric fusion method that clusters intermediate activations of independently trained models into importance-weighted centroids and fits the fused network to them, outperforming baselines in zero-shot non-IID...
-
Assembly of Experts: Linear-time construction of the Chimera LLM variants with emergent and adaptable behaviors
Weight interpolation between two DeepSeek models, restricted to routed expert tensors, yields a 671B Chimera model with near-R1 reasoning and roughly 40% lower output-token count.
-
STAR: Spectral Truncation and Rescale for Model Merging
STAR merges fine-tuned models by truncating small singular values of task vectors and rescaling to restore the nuclear norm, improving multi-task merging performance.
Discussion (0). Continue with ORCID to comment.