Core Concepts
線形近似に二次補正項を追加することで、非線形特徴マップを使用した効率的な補正が可能である。
Abstract
現在の作業は、線形主成分への射影に基づく最良近似が、補正項と組み合わさる際に重要な情報を見逃す可能性があることを示しています。
導入された貪欲法は、特異値に関して降順ではない主成分を選択することを可能にします。
数値実験では、導入された貪欲法により桁違いの高い精度が達成され、百万次元のデータポイントにもスケーリング可能です。
Stats
Numerical experiments demonstrate that an orders of magnitude higher accuracy is achieved with the greedily constructed quadratic manifolds compared to manifolds that are based on the leading principal components alone.
Linear dimensionality reduction in subspaces given by the principal component analysis (PCA) can lead to poor approximations when correlations between components of data points are strongly nonlinear.
The regularization parameter for fitting W is γ = 10^-8 for the greedy approach and also for the approach using the leading r left-singular vectors.
The regularization parameter is set to γ = 10^-2 for alternating minimization, γ = 10^-2 for the quadratic manifold based on the leading r left-singular vectors, and γ = 10^-3 for the proposed greedy approach.
In all examples, all methods have access to the same amount of memory.
Quotes
"Augmenting linear decoder functions with nonlinear correction terms given by feature maps can lead to higher accuracy than linear approximations alone."
"The greedy method introduced in this approach allows selecting principal components that are not necessarily ordered descending with respect to the singular values."
"Numerical experiments demonstrate that an orders of magnitude higher accuracy can be achieved with the introduced greedy method."