In this paper, we propose a novel solution to reduce the computational complexity of convolutional neural network models used for many class image classification. Our proposed model breaks the classification task into three stages: 1) general feature extraction, 2) Mid-level clustering, and 3) hyper-class classification. Steps 1 and 2 could be repeated to build larger hierarchical models. We illustrate that our proposed classifier can reach the level of accuracy reported by the best in class classification models with far less computational complexity (Flop Count) by only activating parts of the model that are needed for the image classification.