Towards Calibrated Deep Clustering Network

Deep clustering has exhibited remarkable performance; however, the <PRE_TAG>overconfidence problem</POST_TAG>, i.e., the estimated confidence for a sample belonging to a particular cluster greatly exceeds its actual prediction accuracy, has been overlooked in prior research. To tackle this critical issue, we pioneer the development of a calibrated <PRE_TAG>deep clustering</POST_TAG> framework. Specifically, we propose a novel <PRE_TAG>dual-head <PRE_TAG><PRE_TAG>deep clustering</POST_TAG> pipeline</POST_TAG></POST_TAG> that can effectively calibrate the estimated confidence and the actual accuracy. The calibration head adjusts the overconfident predictions of the clustering head using <PRE_TAG>regularization methods</POST_TAG>, generating prediction confidence and <PRE_TAG>pseudo-labels</POST_TAG> that match the model learning status. This calibration process also guides the clustering head in dynamically selecting reliable high-confidence samples for training. Additionally, we introduce an effective network initialization strategy that enhances both training speed and network robustness. Extensive experiments demonstrate the proposed calibrated <PRE_TAG>deep clustering</POST_TAG> framework not only surpasses state-of-the-art <PRE_TAG>deep clustering</POST_TAG> methods by approximately 10 times in terms of <PRE_TAG>expected calibration error</POST_TAG> but also significantly outperforms them in terms of <PRE_TAG>clustering accuracy</POST_TAG>.