$C^2M^3$: Cycle-Consistent Multi-Model Merging

TL;DR

This paper introduces a cycle-consistent, data-free method for merging multiple neural network models in weight space, optimizing neuron permutations globally to improve model integration across diverse architectures.

Contribution

It proposes a novel cycle-consistent permutation optimization technique for merging multiple neural networks without data, ensuring error-free composition across models.

Findings

Cycle consistency improves merging accuracy.

Global permutation optimization outperforms local methods.

Activation renormalization enhances merging results.

Abstract

In this paper, we present a novel data-free method for merging neural networks in weight space. Differently from most existing works, our method optimizes for the permutations of network neurons globally across all layers. This allows us to enforce cycle consistency of the permutations when merging $N \geq 3$ models, allowing circular compositions of permutations to be computed without accumulating error along the path. We qualitatively and quantitatively motivate the need for such a constraint, showing its benefits when merging sets of models in scenarios spanning varying architectures and datasets. We finally show that, when coupled with activation renormalization, our approach yields the best results in the task.