Discriminative concept learning network: Reveal high-level differential concepts from shallow architecture

A desired capability of deep learning is to understand the high-level, class-specific features via hierarchical features learning. However the training of deep architectures is costly comparing to simple shallow models. Bringing the high-level feature understanding into a simple shallow architecture remains an open question. We proposed a supervised learning algorithm, enabling binary classification along with an intrinsic ability of learning high-level discriminative concepts via a shallow neural network architecture. The physical architecture of the network has one hidden layer (also serving as the output layer) responsible for the classification and an input layer directly identifies the informative features that constitute the high-level differential concepts between the two classes. Compared to other shallow classifiers, we demonstrate its practicability in real world classification problems. We also illustrate the human-understandable, discriminative concepts learned from the two image recognition exercises. Lastly, we show how it is useful in validating the disease-associated genetic variants in human genome as a real diagnostic genomics application.