Yan Yuxuan ; Lu Na ; Yan Ruofan

Abstract

Combining machine clustering with deep models has shown remarkablesuperiority in deep clustering. It modifies the data processing pipeline intotwo alternating phases: feature clustering and model training. However, suchalternating schedule may lead to instability and computational burden issues.We propose a centerless clustering algorithm called Probability AggregationClustering (PAC) to proactively adapt deep learning technologies, enabling easydeployment in online deep clustering. PAC circumvents the cluster center andaligns the probability space and distribution space by formulating clusteringas an optimization problem with a novel objective function. Based on thecomputation mechanism of the PAC, we propose a general online probabilityaggregation module to perform stable and flexible feature clustering overmini-batch data and further construct a deep visual clustering framework deepPAC (DPAC). Extensive experiments demonstrate that PAC has superior clusteringrobustness and performance and DPAC remarkably outperforms the state-of-the-artdeep clustering methods.