<p>In large-scale machine learning, the computational cost of kernel methods can become prohibitive due to the need to compute pairwise kernel evaluations on extensive datasets. The random feature method is one of the most popular techniques for accelerating kernel methods in large-scale problems while maintaining statistical accuracy. In this paper, we investigate the generalization properties of a robust gradient descent algorithm utilizing random features within a statistical learning framework, where we employ the robust loss function <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(l_{\sigma }\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>l</mi> <mi>σ</mi> </msub> </math></EquationSource> </InlineEquation> instead of the traditional squared loss during training. This loss function is defined by a windowing function <i>G</i> and a scale parameter <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\sigma \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>σ</mi> </math></EquationSource> </InlineEquation>, allowing it to encompass a wide range of commonly used robust losses for regression when <i>G</i> and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\sigma \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>σ</mi> </math></EquationSource> </InlineEquation> are appropriately selected. However, it remains unclear whether the random feature method can preserve statistical accuracy in this context. We analyze the generalization error of the estimator produced by the gradient descent algorithm with random features. Our findings demonstrate that with a suitably chosen scale parameter <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\sigma \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>σ</mi> </math></EquationSource> </InlineEquation> and an appropriate number of random features <i>M</i>, our estimator can converge to the regression function in <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(L^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>L</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>-norm at optimal rates in the mini-max sense (up to a logarithmic term), even if the regression function may not reside in the reproducing kernel Hilbert space.</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

Robust kernel-based gradient descent with random features

  • Qi Hong,
  • Zheng-Chu Guo

摘要

In large-scale machine learning, the computational cost of kernel methods can become prohibitive due to the need to compute pairwise kernel evaluations on extensive datasets. The random feature method is one of the most popular techniques for accelerating kernel methods in large-scale problems while maintaining statistical accuracy. In this paper, we investigate the generalization properties of a robust gradient descent algorithm utilizing random features within a statistical learning framework, where we employ the robust loss function \(l_{\sigma }\) l σ instead of the traditional squared loss during training. This loss function is defined by a windowing function G and a scale parameter \(\sigma \) σ , allowing it to encompass a wide range of commonly used robust losses for regression when G and \(\sigma \) σ are appropriately selected. However, it remains unclear whether the random feature method can preserve statistical accuracy in this context. We analyze the generalization error of the estimator produced by the gradient descent algorithm with random features. Our findings demonstrate that with a suitably chosen scale parameter \(\sigma \) σ and an appropriate number of random features M, our estimator can converge to the regression function in \(L^2\) L 2 -norm at optimal rates in the mini-max sense (up to a logarithmic term), even if the regression function may not reside in the reproducing kernel Hilbert space.