A challenging topic in content-based image retrieval is to determine the discriminant features that improve classification performance. An approach to learn concepts is by estimating mixture model for image databases using EM algorithm; however, this approach is impractical to be implemented for large databases due to the high dimensionality of the feature space. Based on the over-splitting nature of our EM algorithm and the Bayesian analysis of the multiple users' labelling information derived from their relevance feedbacks, we propose a probabilistic MDA to find the discriminating features, and integrate it with the EM framework. The experimental results on Corel images show the effectiveness of concept learning with the probabilistic MDA, and the improvement of the retrieval performance.