Research Papers 论文研究 4h ago Updated 1h ago 更新于 1小时前 42

Scaling Closed-Loop Feature Channel Configuration with LLMs 使用LLM扩展闭环特征通道配置

LLMs can effectively optimize neural network channel configurations via closed-loop code generation and accuracy feedback, with performance improvements scaling in denser sampling regimes. The best model achieved 36.76% accuracy on CIFAR-100 using only 11.8M parameters, a significant improvement in parameter efficiency over earlier models that required 166.5M parameters for lower accuracy. Analysis of 462 verified candidates revealed architectural regularities, such as non-power-of-two channel w 研究验证了基于LLM的闭环特征通道配置搜索在密集采样 regime 下的可扩展性,证明优化行为具有鲁棒性。 在CIFAR-100数据集上,经过8个周期的迭代,最佳模型准确率从0.3144提升至0.3676,且参数效率显著提高(11.8M vs 166.5M参数)。 大规模样本揭示了新的架构规律:非2的幂次通道宽度占41.8%,高性能模型倾向于“中等早期宽度+扩展中后期块”的结构。 该方法通过可执行代码生成与准确率反馈直接优化神经网络宽度,无需传统神经架构搜索(NAS)的复杂奖励机制。

55
Hot 热度
65
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • LLMs can effectively optimize neural network channel configurations via closed-loop code generation and accuracy feedback, with performance improvements scaling in denser sampling regimes.
  • The best model achieved 36.76% accuracy on CIFAR-100 using only 11.8M parameters, a significant improvement in parameter efficiency over earlier models that required 166.5M parameters for lower accuracy.
  • Analysis of 462 verified candidates revealed architectural regularities, such as non-power-of-two channel widths occurring in 41.8% of cases and a preference for moderate early widths with expanded middle or later blocks in high-performing models.

Why It Matters

This research demonstrates that Large Language Models can serve as effective agents for automated neural architecture search (NAS), specifically for optimizing internal layer structures like channel widths. It provides empirical evidence that LLM-generated architectures can surpass human-designed or traditionally searched counterparts in terms of parameter efficiency, offering a new pathway for lightweight model design.

Technical Details

  • Methodology: A closed-loop search process where LLMs generate executable code for neural networks, which are then evaluated for accuracy on CIFAR-100, providing feedback for subsequent iterations.
  • Scale: The study scaled the search to 250 candidate networks per fine-tuning cycle, analyzing 2000 generated candidates across 8 cycles, resulting in 462 verified evaluations after filtering.
  • Performance Trends: Per-cycle mean accuracy showed a positive linear trend (slope 9.87e-4, p=0.043). The high-performing frontier improved from 31.44% to 36.76% accuracy.
  • Architectural Findings: High-performing models exhibited structured channel-allocation patterns, typically featuring moderate channel widths in early layers and expanded widths in middle or later blocks. Non-power-of-two widths were common (41.8%).

Industry Insight

  • Efficiency Gains: Practitioners should consider LLM-driven NAS for scenarios requiring high parameter efficiency, as the results show substantial reductions in model size without sacrificing accuracy.
  • Design Heuristics: The identified regularities (moderate early widths, expanded later blocks) provide actionable heuristics for manual architecture design or as constraints for future automated search algorithms.
  • Scalability of LLM-NAS: The transfer of optimization signals to denser sampling regimes suggests that LLM-based architecture search is robust and scalable, warranting further investment in integrating LLMs into the model development pipeline.

TL;DR

  • 研究验证了基于LLM的闭环特征通道配置搜索在密集采样 regime 下的可扩展性,证明优化行为具有鲁棒性。
  • 在CIFAR-100数据集上,经过8个周期的迭代,最佳模型准确率从0.3144提升至0.3676,且参数效率显著提高(11.8M vs 166.5M参数)。
  • 大规模样本揭示了新的架构规律:非2的幂次通道宽度占41.8%,高性能模型倾向于“中等早期宽度+扩展中后期块”的结构。
  • 该方法通过可执行代码生成与准确率反馈直接优化神经网络宽度,无需传统神经架构搜索(NAS)的复杂奖励机制。

为什么值得看

本文提供了LLM作为神经架构搜索代理的实证证据,表明其不仅能生成有效代码,还能通过闭环反馈实现性能与效率的双重优化。对于希望探索自动化模型设计或降低计算成本的AI从业者而言,该研究展示了LLM在微观架构调整中的巨大潜力和具体收益。

技术解析

  • 实验规模与设置:将搜索范围扩展至每轮微调周期250个候选网络,共进行8个完整周期,生成2000个候选者,经筛选后获得462个有效的CIFAR-100评估结果,确保了统计显著性。
  • 性能提升趋势:每轮平均准确率呈现正线性增长趋势(斜率9.87e-4, p=0.043),前沿性能(Top-5/Top-10均值)改善更为明显,证明了LLM在持续迭代中能捕捉到有效的优化信号。
  • 参数效率突破:最终最佳模型仅用11.8M参数即达到0.3676准确率,而早期高性能模型需166.5M参数,显示出LLM生成的架构在稀疏性和效率上的巨大优势。
  • 架构规律发现:数据分析显示41.8%的验证候选者使用非2的幂次通道宽度;最强模型普遍具备结构化通道分配模式,即早期层保持适度宽度,而中间或后期层显著扩展。

行业启示

  • LLM-NAS范式验证:证实了利用LLM生成可执行代码并结合反馈进行闭环优化的可行性,为未来开发更轻量级、无需梯度反向传播的自动化架构搜索工具提供了理论支持。
  • 打破传统结构偏见:研究结果挑战了传统CNN设计中严格遵循2的幂次通道数的惯例,提示业界应更加开放地探索非标准通道宽度以换取更高的参数效率和性能。
  • 关注架构归纳偏置:LLM在优化过程中自发涌现出的“前窄后宽”或特定块扩展模式,可能隐含了对数据特征提取难度的深刻理解,值得进一步研究这些规律是否可迁移至其他视觉任务。

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。

LLM 大模型 Code Generation 代码生成 Research 科学研究