LGMLNov 11, 2013

Learning Mixtures of Linear Classifiers

arXiv:1311.2547v423 citations
Originality Incremental advance
AI Analysis

This addresses a discriminative learning problem for machine learning practitioners, offering a provably efficient alternative to existing methods like EM.

The paper tackles the problem of learning mixtures of linear classifiers without provable guarantees, developing a spectral method with a mirroring trick that achieves nearly optimal statistical efficiency under a probabilistic assumption on feature distribution.

We consider a discriminative learning (regression) problem, whereby the regression function is a convex combination of k linear classifiers. Existing approaches are based on the EM algorithm, or similar techniques, without provable guarantees. We develop a simple method based on spectral techniques and a `mirroring' trick, that discovers the subspace spanned by the classifiers' parameter vectors. Under a probabilistic assumption on the feature vector distribution, we prove that this approach has nearly optimal statistical efficiency.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes