We present a tree-structured network architecture for large-scale image classification. The trunk of the network contains convolutional layers optimized over all classes. At a given depth, the trunk splits into separate branches, each dedicated to discriminate a different subset of classes. Each branch acts as an expert classifying a set of categories that are difficult to tell apart, while the trunk provides common knowledge to all experts in the form of shared features. The training of our "network of experts" is completely end-to-end: the partition of categories into disjoint subsets is learned simultaneously with the parameters of the network trunk and the experts are trained jointly by minimizing a single learning objective over all classes. The proposed structure can be built from any existing convolutional neural network (CNN).
Accuracy on CIFAR-100 dataset. We achieve the best accuracy result (76.24) using ResNet56 as a base model.
|Architecture||Base Model %||NofE (Our Network of Experts) %|
Top-1 accuracy on the ImageNet validation set using AlexNet and our NoFE.
|NofE, K=10 (fully-balanced)||61.29|
|NofE, K=40 (fully-balanced)||60.85|
The trained NoFE models can be downloaded from the following links:
AlexNet-C100: NofE model based on AlexNet architecture trained on Cifar100. [Download]
AlexNet-Quick-C100: NofE model based on AlexNet-Quick architecture trained on Cifar100. [Download]
VGG-C100: NofE model based on VGG architecture trained on Cifar100. [Download]
NIN-C100: NofE model based on Network in Network architecture trained on Cifar100. [Download]
ResNet56-C100: NofE model based on ResNet 56 layers architecture trained on Cifar100. [Download]
AlexNet-ImageNet: NofE model based on AlexNet architecture trained on ImageNet. [Download]
Karim Ahmed, Mohammad Haris Baig, Lorenzo Torresani
Network of Experts for Large-Scale Image Categorization
This material is based upon work supported by the National Science Foundation under CAREER award IIS-0952943. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation (NSF).