The Empirical Impact of Reducing Symmetries on the Performance of Deep Ensembles and MoE

2025-02-24

AI Summary

This work investigates the impact of neural network symmetry on the performance of Deep Ensembles (DE) and Mixture of Experts (MoE). For DE, the authors systematically validate that symmetry-breaking training yields consistent gains, providing the first empirical evidence that asymmetric DE significantly outperforms its symmetric counterpart as model scale increases. For MoE, they find that reducing symmetry confers negligible benefits; consequently, they propose Mixture of Interpolated Experts (MoIE), a novel architecture that explicitly incorporates linear mode connectivity into expert models, thereby extending the applicability of linear interpolation in MoE. Methodologically, the study integrates symmetry-breaking training, linear mode connectivity analysis, and cross-task evaluation across multiple datasets. Key contributions are: (1) establishing the universality of asymmetric architectures for improving DE performance; and (2) introducing MoIE and empirically demonstrating its effectiveness in enhancing mode connectivity.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Computer Vision: Large Vision Models
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Recent studies have shown that reducing symmetries in neural networks enhances linear mode connectivity between networks without requiring parameter space alignment, leading to improved performance in linearly interpolated neural networks. However, in practical applications, neural network interpolation is rarely used; instead, ensembles of networks are more common. In this paper, we empirically investigate the impact of reducing symmetries on the performance of deep ensembles and Mixture of Experts (MoE) across five datasets. Additionally, to explore deeper linear mode connectivity, we introduce the Mixture of Interpolated Experts (MoIE). Our results show that deep ensembles built on asymmetric neural networks achieve significantly better performance as ensemble size increases compared to their symmetric counterparts. In contrast, our experiments do not provide conclusive evidence on whether reducing symmetries affects both MoE and MoIE architectures.

Problem

• Impact of symmetry reduction
• Deep ensembles performance
• Mixture of Experts effectiveness

Innovation

• Reduced symmetries enhance connectivity
• Asymmetric networks improve ensemble performance
• Mixture of Interpolated Experts introduced