Learning Visual Models using a Knowledge Graph as a Trainer

원문 초록

Traditional computer vision approaches, based on neural networks (NN), are typically trained on a large amount of image data. By minimizing the cross-entropy loss between a prediction and a given class label, the NN and its visual embedding space are learned to fulfill a given task. However, due to the sole dependence on the image data distribution of the training domain, these models tend to fail when applied to a target domain that differs from their source domain. To learn a more robust NN to domain shifts, we propose the knowledge graph neural network (KG-NN), a neuro-symbolic approach that supervises the training using image-data-invariant auxiliary knowledge. The auxiliary knowledge is first encoded in a knowledge graph with respective concepts and their relationships, which is then transformed into a dense vector representation via an embedding method. Using a contrastive loss function, KG-NN learns to adapt its visual embedding space and thus its weights according to the image-data invariant knowledge graph embedding space. We evaluate KG-NN on visual transfer learning tasks for classification using the mini-ImageNet dataset and its derivatives, as well as road sign recognition datasets from Germany and China. The results show that a visual model trained with a knowledge graph as a trainer outperforms a model trained with cross-entropy in all experiments, in particular when the domain gap increases. Besides better performance and stronger robustness to domain shifts, these KG-NN adapts to multiple datasets and classes without suffering heavily from catastrophic forgetting.