Fast decentralized learning of a Gaussian mixture model for large-scale multimedia retrieval

We address herein the distributed computation of a probability density estimate. Class-conditional probability density estimation is a central need in multimedia pattern recognition, but has classically be conducted in a centralized fashion. In contrast, the present work is motivated by the perspective of a multimedia indexing and retrieval peer-to-peer system over the Internet. In a decentralized fashion, algorithms and data from various contributors would cooperate towards a collective statistical learning. A typical need is aggregation of probabilistic Gaussian mixture models describing the same class, but estimated on several nodes on different data sets. We tackle this goal through an approach requiring only moderate computation at each node and little data to transit between nodes. Both properties are obtained by fusion models via their (few) parameters, rather than via multimedia data itself. Estimation of the aggregated model is provided by an iterative scheme, derived from a modification on Kullback divergence. We provide experimental results on a speaker recognition task with real data, in a gossip propagation setting.