Low-cost distributed learning of a Gaussian mixture model for multimedia content-based indexing on a peer-to-peer network

This work deals with estimation of a probability density, which is a common issue in multimedia pattern recognition. The originality comes from its computation in a distributed manner, since the study 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.In this setting, aggregation of probabilistic Gaussian mixture models of the same class, but estimated on several nodes on different data sets, is a typical need, which we address herein. The proposed approach for fusion only requires moderate computation at each node and little data to transit between nodes. Both properties are obtained by aggregating models via their (few) parameters, rather than via multimedia data itself. Mixture models are in fact concatenated, then reduced to a suitable number of Gaussian components. A modification on Kullback divergence leads to an iterative scheme for estimating this aggregated model. We provide experimental results on a speaker recognition task with real data, in a gossip propagation setting.