Hierarchical Softmax for Fine-Grained Classification

Abstract

Normally, we use Softmax after the last fully connected layer in a convolutional neural network to get the probability of the classes. With the number of increasing classes, the accuracy for classification decreases. Also, Multiple datasets cannot be jointly trained because they are not mutually exclusive. A joint training algorithm that combines multiple datasets in a hierarchical structure and uses hierarchical softmax as the output layer is proposed. We use this technique to train the detection and classification datasets like COCO and ImageNet together. We train YOLO on 9000 classes from the detection dataset of COCO and classification dataset of ImageNet. Our trained model was able to predict classes that were not in the training dataset. While testing on the 200 detection classes of ImageNet, we were able to achieve 19 % mean Average Precision. Out of this 200 classes the model was trained only on 44 classes that were present in the COCO detection dataset. We achieved 16 % mAP on the 156 classes that were never in the training dataset. Apart from this, Our model was trained to predict over 9000 classes.