CEN: Concept Evolution Network for Image Classification Tasks

Abstract

Image classification is a challenging but fundamental task for many computer vision applications, such as self-driving, face recognition, and object tracking. The deep neural network (DNN) is a modern, powerful model to tackle this task, whose representation ability mainly comes from hidden layers. The interpretability of DNN, however, drops rapidly as the inexplicable hidden part becomes deeper and deeper. To make neural networks more explainable, we propose a novel neural network named concept evolution network (CEN), learning explicit concepts of images to help classify. Concepts evolve during training with three stages: emergence, elevation, and elimination. We design three algorithms (one primary and two improved) to train CEN. The experiment results on MNIST show our methods' feasibility and that CEN has both interpretability and adaptive learning capacity for the image classification task. In the last section, we discuss the development prospects of CEN in the future.