A compositional parts based model for object recognition

Deep Learning architectures based on Convolutional Neural Networks (CNNs) are the state of the art technique for object recognition and segmentation. However, CNNs require millions of labelled images to reach this performance. They also suffer from a lack of interpretability making it difficult to understand why some images are misclassified. Recent work on adversarial attacks on these types of systems has highlighted this problem. Alternatives to CNNs such as parts based compositional models rivalled or surpassed CNNs performance in the past before the advances in network architecture and large-scale GPU training allowed CNNs to reach their current performance. Parts based compositional models can overcome some of the weaknesses of CNNs, but themselves suffer from poor scalability on large-scale datasets. A compositional model built using neural network components, but forming an interpretable compositional graph would in principle combine the advantages of both approaches in a scaleable, interpretable model.