Image classification using the conventional cross-entropy loss is a common approach in the field of computer vision. Although with the available resources it has been able to achieve high prediction accuracy, the results are not intuitive especially when they are misclassified. In cases, such as image retrieval and autonomous driving, retrieving meaningful or related classes and reducing the extent of mistake made are necessary, respectively. Thus, we introduce a method to learn the semantic relationship among the classes while training a model that forms a feature space of similar classes grouped together. Our results are comparable with the traditional training methods and reduce the extent of misclassification by a model.

Convolutional neural networks (ConvNets) have become increasingly popular for image classification tasks. All contemporary computer vision problems are being dominated by ConvNets. Conventional training methods using cross-entropy loss for training have constantly outperformed the state-of-the-art technique to set a new standard in the ImageNet classification challenge. However, growing accuracy come at the cost of enormous number of parameters and computations. Further, classical learning algorithms do not utilize the semantic relationship between the classes present in the dataset. Thus, interpreting the behavior of the model become difficult even though the results may be desirable. Hence, we demonstrate a classification method by leveraging the WordNet hierarchy on the ImageNet dataset to establish class relationships and label embedding. The model is trained using cross entropy with soft labels based on the semantic similarity between the generated output and the ground truth. Unlike categorical cross entropy, it does not treat every predicted label as equally erroneous. The method generates meaningful neighboring classes in the feature space of the true label.