Half Quadratic Dual Learning for Fuzzy Multiconcepts of Partially-Observed Images

For partially-observed data with large continuous occlusion or specific outlier noise, it is difficult to represent them by using L
₁
-, L
₂
-, and L
_2,1
-norms, not to mention learning multiconcept mapping from them. Furthermore, multiconcept learning (i.e., multilabel learning) and sample-correlation learning are inseparable as they provide guiding information to each other. Nonetheless, when most sample relations are unknown/biased, detecting multiconcepts associated with partially-observed data becomes even more difficult. In view of such challenges, this study presents a dual learning model that can simultaneously learn multiconcepts, sample relations, and data representation while partially-observed data and partially-observed sample relations are present. Moreover, half quadratic (HQ) loss functions along with half quadratic optimization and their gradient forms are developed in this study to model unobserved information. Different types of partially-observed sources are jointly hinged together to compensate each other during learning. This allows the system to estimate accurate multiconcepts, sample relations, and data representation. Finally, fuzzy scores for multiconcepts can be learnt and generated. Thus, a labelled data instance can utilize such fuzzy scores to rank its concepts. Experiments on open images datasets were carried out for evaluating the performance. Analytical results showed that the proposed model was capable of conquering multiconcept learning problems while data were partially observed, better than the baselines, thereby demonstrating the effectiveness of the proposed idea.