Patch Affinity Propagation

Affinity propagation constitutes an exemplar based clustering technique which reliably optimizes the quantization error given a matrix of pairwise data dissimilarities by means of the max-sum algorithm for factor graphs. Albeit very efficient for sparse matrices, it displays squared complexity in the worst case, hence it is not suited as high throughput method due to time and memory constraints. We propose an extension of affinity propagation to patch clustering such that data are treated in chunks of fixed size with limited memory requirements and linear time. We test the suitability of the approach for two biomedical applications.