Dimensionality reduction in decentralized networks by Gossip aggregation of principal components analyzers

Jérôme Fellus (1) , David Picard (1) , Philippe-Henri Gosselin (2, 3)

1 MIDI - Multimedia Indexation and Data Integration
2 ETIS - UMR 8051 - Equipes Traitement de l'Information et Systèmes
3 TEXMEX - Multimedia content-based indexing

This paper considers dimensionality reduction in large decentralized networks with limited node-local computing and memory resources and unreliable point-to-point connectivity (e.g peer-to-peer, sensors or ad-hoc mobile networks). We propose an asynchronous decentralized algorithm built on a Gossip consensus protocol that perform Principal Components Analysis (PCA) of data spread over such networks. All nodes obtain the same local basis that span the global principal subspace. Reported experiments show that obtained bases both reach a consensus and accurately estimate the global PCA solution.

Mots clés: PCA; Gossip; Learning; Distributed

Domaines: Machine Learning [stat.ML]

Jérôme Fellus, David Picard, Philippe-Henri Gosselin. Dimensionality reduction in decentralized networks by Gossip aggregation of principal components analyzers. European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Apr 2014, Bruges, Belgium. pp.171-176. ⟨hal-00985721⟩