Research Papers 论文研究 5h ago Updated 1h ago 更新于 1小时前 45

When Muon Meets Task Interference: A Spectral Perspective on Continual Learning and Model Merging 当Muon遇见任务干扰:持续学习与模型合并的光谱视角

Continual learning (catastrophic forgetting) and model merging (weight-disentanglement error) are unified under a single concept: task interference, where a parameter update for one task shifts outputs on another. Task interference is formalized as a layer-wise Frobenius inner product ⟨ΔW_ℓ, J_ℓ(x)⟩_F, with a theoretical upper bound isolating the spectral norm ||ΔW_ℓ||_2 as an optimizer-controllable factor. The Muon optimizer is identified as a mechanism that intrinsically regulates spectral nor 持续学习(CL)和模型合并(MM)面临的灾难性遗忘与权重解耦误差本质上是同一现象:任务干扰(task interference) 作者将任务干扰形式化为层级Frobenius内积⟨ΔW_ℓ, J_ℓ(x)⟩_F,并推导出谱范数‖ΔW_ℓ‖₂作为优化器可控的上界因子 Muon优化器通过构造天然控制谱范数,从而收紧任务干扰上界,为CL和MM提供统一的优化器视角解决方案 实验验证:在8任务模型合并基准上,Muon替换AdamW提升高达+5.02分;在持续学习中覆盖10个类增量协议、3个任务增量协议及11任务MTIL基准均获正向收益

55
Hot 热度
75
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Continual learning (catastrophic forgetting) and model merging (weight-disentanglement error) are unified under a single concept: task interference, where a parameter update for one task shifts outputs on another.
  • Task interference is formalized as a layer-wise Frobenius inner product ⟨ΔW_ℓ, J_ℓ(x)⟩_F, with a theoretical upper bound isolating the spectral norm ||ΔW_ℓ||_2 as an optimizer-controllable factor.
  • The Muon optimizer is identified as a mechanism that intrinsically regulates spectral norm, thereby tightening the task interference bound for both CL and MM.
  • Replacing AdamW with Muon yields up to +5.02 accuracy points on an eight-task model-merging benchmark across three CLIP backbones.
  • Muon delivers uniformly positive gains across ten class-incremental protocols, three task-incremental protocols, and the 11-task MTIL benchmark for continual learning.

Why It Matters

This work reframes two major challenges in multi-task AI—catastrophic forgetting and model merging incompatibility—as manifestations of the same underlying phenomenon, offering a unified theoretical lens. For practitioners, it demonstrates that optimizer choice is not merely an implementation detail but a principled lever for controlling task interference, with the Muon optimizer providing measurable gains across diverse benchmarks.

Technical Details

  • Task Interference Formalization: The paper defines task interference as a layer-wise Frobenius inner product between the weight update ΔW_ℓ and the Jacobian J_ℓ(x) of layer outputs, capturing how updates for one task perturb performance on another.
  • Spectral Norm Bound: A theoretical upper bound is derived that isolates the spectral norm ||ΔW_ℓ||_2 as the key optimizer-controllable factor governing task interference; per-mode analysis confirms this bound tracks the dominant empirical interference.
  • Muon as Spectral Regulator: Muon is shown to regulate spectral norm by construction, making it a principled optimizer-centric solution that tightens the interference bound without requiring task-specific modifications.
  • Experimental Validation: Benchmarks include an eight-task model-merging setup across three CLIP backbones (up to +5.02 accuracy over AdamW), plus continual learning evaluations across ten class-incremental protocols, three task-incremental protocols, and the 11-task MTIL benchmark with uniformly positive gains.

Industry Insight

  • Optimizer selection should be treated as a strategic decision in multi-task and continual learning pipelines, not a hyperparameter afterthought; Muon's spectral norm control offers a plug-and-play improvement over AdamW.
  • The unified task interference framework suggests that future research on model merging and catastrophic forgetting can share theoretical tools and solutions, accelerating progress across both subfields.
  • For production systems deploying merged models or incremental learning, adopting Muon could yield immediate accuracy gains with minimal engineering overhead, particularly in vision-language settings like CLIP.

TL;DR

  • 持续学习(CL)和模型合并(MM)面临的灾难性遗忘与权重解耦误差本质上是同一现象:任务干扰(task interference)
  • 作者将任务干扰形式化为层级Frobenius内积⟨ΔW_ℓ, J_ℓ(x)⟩_F,并推导出谱范数‖ΔW_ℓ‖₂作为优化器可控的上界因子
  • Muon优化器通过构造天然控制谱范数,从而收紧任务干扰上界,为CL和MM提供统一的优化器视角解决方案
  • 实验验证:在8任务模型合并基准上,Muon替换AdamW提升高达+5.02分;在持续学习中覆盖10个类增量协议、3个任务增量协议及11任务MTIL基准均获正向收益

为什么值得看

本文首次从谱范数角度统一解释持续学习与模型合并的核心挑战,揭示两者本质同源,为优化器设计提供了新的理论框架。Muon优化器作为新兴方案被证明具有 principled 的理论优势,对实际工程选型具有重要参考价值。

技术解析

  • 理论框架:将任务干扰定义为参数更新对非目标任务输出的偏移,形式化为层级的Frobenius内积⟨ΔW_ℓ, J_ℓ(x)⟩_F,其中J_ℓ(x)为第ℓ层雅可比矩阵
  • 谱范数上界:推导出任务干扰的理论上界,分离出谱范数‖ΔW_ℓ‖₂作为关键可控因子,per-mode分析表明该上界能追踪经验干扰的主要部分
  • Muon优化器机制:Muon通过正交变换和符号矩阵乘法构造更新方向,天然具有控制谱范数的性质,无需额外正则化即可收紧干扰上界
  • 实验设置:模型合并基准使用三个CLIP backbone在8任务上测试;持续学习覆盖10个类增量协议、3个任务增量协议及11任务MTIL基准
  • 性能提升:Muon替换AdamW在模型合并任务上最高提升+5.02分,在持续学习各协议中均呈现一致正向增益

行业启示

  • 优化器选择不应仅视为工程细节,其诱导的几何结构直接影响多任务学习中的干扰控制,建议在CL和MM场景中优先评估Muon等谱范数可控的优化器
  • 持续学习与模型合并的解决方案可借鉴统一理论框架,现有方法可基于任务干扰视角重新审视与改进
  • 谱范数控制作为优化器设计的理论指导原则,有望扩展至其他多任务学习场景,如多任务训练、模型编辑等

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Training 训练 LLM 大模型