Scholarlink

<p>Liquid Neural Networks (LNNs), characterized by their continuous-time dynamical formulation, offer strong performance in time-series forecasting but suffer from high inference costs due to reliance on numerical Ordinary Differential Equation (ODE) solvers. To mitigate this limitation, this paper introduces an efficient LNN architecture incorporating a conditional skip mechanism that exploits the temporally non-uniform structure of cloud workloads. A parallel framework is constructed with a precise path preserving the original ODE-based computation and an approximate path employing parameterized linear interpolation. A skip decision module, informed by local statistical measures and combined with adaptive step-size control, enables dynamic regulation of computational density. Theoretical analysis shows that under a bounded skip policy, prediction error remains controlled and scales linearly with the product of the skip rate and step size, yielding a complexity reduction of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text{O}(\text{d})\)</EquationSource> </InlineEquation>. Empirical results on real-world cloud workload datasets validate that the proposed method enables highly efficient inference acceleration while maintaining comparable accuracy, demonstrating its practicality for deployment in resource-constrained environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Conditional skip liquid neural networks: an efficient inference framework for large-scale time-series cloud resource prediction

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Enliang Wang,
  • Jiajun Wang,
  • Xintong Wang,
  • Zhixin Sun

摘要

Liquid Neural Networks (LNNs), characterized by their continuous-time dynamical formulation, offer strong performance in time-series forecasting but suffer from high inference costs due to reliance on numerical Ordinary Differential Equation (ODE) solvers. To mitigate this limitation, this paper introduces an efficient LNN architecture incorporating a conditional skip mechanism that exploits the temporally non-uniform structure of cloud workloads. A parallel framework is constructed with a precise path preserving the original ODE-based computation and an approximate path employing parameterized linear interpolation. A skip decision module, informed by local statistical measures and combined with adaptive step-size control, enables dynamic regulation of computational density. Theoretical analysis shows that under a bounded skip policy, prediction error remains controlled and scales linearly with the product of the skip rate and step size, yielding a complexity reduction of \(\text{O}(\text{d})\) . Empirical results on real-world cloud workload datasets validate that the proposed method enables highly efficient inference acceleration while maintaining comparable accuracy, demonstrating its practicality for deployment in resource-constrained environments.