Research Papers 论文研究 3h ago Updated 1h ago 更新于 1小时前 49

Agentic Neural Architecture Search 代理式神经架构搜索

Introduces AgentNAS, a hybrid framework combining LLM-driven design with traditional Neural Architecture Search (NAS) to automate architecture discovery without manual search space engineering. Utilizes a "slotted architecture" concept where an LLM generates a seed design that is decomposed into interchangeable modules, creating a bounded, task-specific search space for NAS optimization. Achieves state-of-the-art results on 11 out of 17 diverse tasks (classification, regression, segmentation, ta 提出AgentNAS框架,利用LLM生成种子架构并分解为“插槽化架构”,自动定义无需人工干预的有界搜索空间。 在17个跨模态任务(分类、回归、分割等)中,于11个任务上取得SOTA,超越包括专家设计在内的基线模型。 验证了LLM设计与NAS搜索的互补性:LLM提供高质量初始结构,NAS通过插槽组合优化带来额外性能增益。 该方法在不同能力的LLM上均表现稳健,证明了人机分工机制在神经架构搜索中的有效性。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces AgentNAS, a hybrid framework combining LLM-driven design with traditional Neural Architecture Search (NAS) to automate architecture discovery without manual search space engineering.
  • Utilizes a "slotted architecture" concept where an LLM generates a seed design that is decomposed into interchangeable modules, creating a bounded, task-specific search space for NAS optimization.
  • Achieves state-of-the-art results on 11 out of 17 diverse tasks (classification, regression, segmentation, tagging) across NAS-Bench-360 and Unseen NAS datasets.
  • Demonstrates that LLM-generated seeds and NAS combinatorial recombination are complementary, with the LLM providing strong initial performance and NAS delivering further gains through slot-based optimization.
  • Validates the robustness of this division of labor across three different LLMs of varying capability levels, confirming the generalizability of the approach.

Why It Matters

This research addresses a critical bottleneck in automated machine learning: the reliance on manually engineered search spaces that require significant domain expertise. By leveraging LLMs to create flexible, slotted scaffolds, AgentNAS democratizes access to high-performance architecture design, allowing practitioners to apply advanced NAS techniques to new tasks without prior architectural knowledge. This synergy between generative AI and optimization algorithms sets a new standard for efficiency and adaptability in neural architecture search.

Technical Details

  • Slotted Architecture Mechanism: The core innovation involves an LLM generating a seed architecture which is then parsed into a scaffold with named, interchangeable module slots. This transforms the open-ended LLM output into a structured, bounded search space suitable for conventional NAS algorithms.
  • Three-Phase Pipeline (AgentNAS): The system operates in distinct phases: (1) LLM generation of the seed architecture, (2) decomposition into slotted structures, and (3) NAS-driven exploration and optimization within the defined slots. This modularity allows for independent measurement of each component's contribution.
  • Comprehensive Benchmarking: Evaluated on 17 tasks spanning multiple modalities and problem types, including image classification, dense regression, semantic segmentation, and multi-label tagging. The evaluation utilized both established benchmarks (NAS-Bench-360) and unseen tasks to test generalization.
  • Complementary Search Strategies: Ablation studies highlight that while LLM seeds alone often outperform existing baselines, the addition of NAS provides significant incremental improvements through combinatorial recombination of slots, a process difficult for LLMs to achieve independently.

Industry Insight

  • Shift from Manual Design to Hybrid Automation: Organizations should move away from relying solely on human experts for architecture design or purely generative LLM approaches. Integrating LLMs for initial structure generation with rigorous NAS for optimization offers a superior balance of creativity and precision.
  • Reduced Barrier to Entry for Advanced ML: The ability to automatically define task-specific search spaces means that teams lacking deep NAS expertise can still leverage advanced architecture search techniques, accelerating R&D cycles for specialized models.
  • Robustness Across Model Capabilities: Since the method performs consistently across different LLM tiers, companies can optimize costs by using smaller or less expensive LLMs for the seed generation phase while reserving computational resources for the NAS optimization phase, ensuring scalability.

TL;DR

  • 提出AgentNAS框架,利用LLM生成种子架构并分解为“插槽化架构”,自动定义无需人工干预的有界搜索空间。
  • 在17个跨模态任务(分类、回归、分割等)中,于11个任务上取得SOTA,超越包括专家设计在内的基线模型。
  • 验证了LLM设计与NAS搜索的互补性:LLM提供高质量初始结构,NAS通过插槽组合优化带来额外性能增益。
  • 该方法在不同能力的LLM上均表现稳健,证明了人机分工机制在神经架构搜索中的有效性。

为什么值得看

本文解决了传统NAS依赖人工设计搜索空间且难以泛化的痛点,同时克服了纯LLM生成架构缺乏系统性优化的局限。它为自动化机器学习(AutoML)提供了新的范式,展示了如何结合大模型的创造力与专用搜索算法的精确性,对降低模型开发门槛具有重要意义。

技术解析

  • 核心机制:引入“插槽化架构”(Slotted Architecture)概念,将LLM生成的完整架构分解为带有命名、可互换模块插槽的脚手架,从而自动生成针对特定任务的有界搜索空间。
  • Pipeline架构:AgentNAS采用模块化三阶段流水线,每个组件的贡献可独立测量,实现了LLM驱动的设计与NAS驱动的搜索之间的明确分工。
  • 实验验证:在NAS-Bench-360和Unseen NAS数据集上进行了广泛评估,涵盖分类、密集回归、分割和多标签标记等多种任务类型。
  • 消融研究结论:LLM生成的种子架构本身已优于多数基线,而NAS通过跨插槽的组合重排进一步提升了性能,这种组合搜索模式是独立LLM采样无法复制的。

行业启示

  • 混合智能成为趋势:未来的AI系统开发将更多依赖“大模型创意+专用算法优化”的混合模式,而非单一依赖某一种技术路径。
  • 降低AutoML门槛:通过自动化搜索空间的构建,可以大幅减少对领域专家的依赖,使神经架构搜索更易于应用于新兴或小众任务。
  • 模块化设计价值:将架构生成与搜索过程解耦为独立模块,不仅提高了系统的可解释性和灵活性,也为后续针对不同LLM或NAS算法的替换提供了标准化接口。

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 Agent Agent LLM 大模型 Training 训练