Core Concepts
Local minima in Gaussian Mixture Models share common structures identifying true cluster centers.
Abstract
Investigates negative log-likelihood landscape of GMMs.
Reveals common structure of local minima.
Two types of sub-configurations identified.
Fine-grained analysis for one-dimensional GMMs provided.
Theoretical results applicable to over/under-specified components.
Stats
負の対数尤度関数の局所最小値を特定する条件が提供されます。
すべての局所最小値は真のクラスタ中心を部分的に識別する共通構造を共有します。
Quotes
"Each local minimum involves two types of sub-configurations."
"All local minima partially identify the means of the true mixture model."