Hierarchical Classification for Large-Scale Learning

Deep neural networks (DNNs) have drawn much attention due to their success in various vision tasks. Current DNNs are used on data with a relatively small number of classes (e.g., 1000 or less) and employ a fully connected layer for classification, which allocates one neuron for each class and thus, per-example, the classification scales as O(K) with the number of classes K. This approach is computationally intensive for many real-life applications where the number of classes is very large (e.g., tens of thousands of classes). To address this problem, our paper introduces a hierarchical approach for classification with a large number of classes that scales as O(K) and could be extended to O(logK) with a deeper hierarchy. The method, called Hierarchical PPCA, uses a self-supervised pretrained feature extractor to obtain meaningful features and trains Probabilistic PCA models on the extracted features for each class separately, making it easy to add classes without retraining the whole model. The Mahalanobis distance is used to obtain the classification result. To speed-up classification, the proposed Hierarchical PPCA framework clusters the image class models, represented as Gaussians, into a smaller number of super-classes using a modified k-means clustering algorithm. The classification speed increase is obtained by Hierarchical PPCA assigning a sample to a small number of the most likely super-classes and restricting the image classification to the image classes corresponding to these super-classes. The fact that the model is trained on each class separately makes it applicable to training on very large datasets such as the whole ImageNet with more than 10,000 classes. Experiments on three standard datasets (ImageNet-100, ImageNet-1k,and ImageNet-10k) indicate that the hierarchical classifier can achieve a superior accuracy with up to a 16-fold speed increase compared to a standard fully connected classifier.