A Discriminative Vectorial Framework for Multi-Modal Feature Representation

Due to the rapid advancements of sensory and computing technology, multi-modal data sources that represent same pattern or phenomenon have attracted growing attention. As a result, finding means explore useful information from these has quickly become necessity. In this paper, discriminative vectorial framework is proposed for feature representation in knowledge discovery by employing hashing (MH) correlation maximization (DCM) analysis. Specifically, capable minimizing semantic similarity among different modalities MH exacting intrinsic representations across multiple DCM analysis jointly, enabling novel representation. Moreover, strategy analyzed further optimized based on canonical non-canonical cases, respectively. Consequently, generated leads effective utilization input high quality, producing improved, sometimes quite impressive, results various applications. The effectiveness generality are demonstrated utilizing classical features deep neural network (DNN) with applications image multimedia recognition tasks, including visualization, face recognition, object recognition; cross-modal (text-image) audio emotion recognition. Experimental show solutions superior state-of-the-art statistical machine learning (SML) DNN algorithms.