Optimized Visual Recognition Through a Deep Convolutional Neural Network With Hierarchical Modular Organization

Artificial Intelligence, including machine learning and deep convolutional neural networks (DCNNs), relies on complex algorithms and neural networks to process and analyze data. DCNNs for visual recognition often require access to high-performance hardware, such as GPUs and cloud-based computing resources, to perform tasks efficiently. Visual recognition requires DCNNs with numerous layers. Fully connected layers in DCNNs are often the most computationally intensive. These layers connect every neuron in one layer to every neuron in the next layer, resulting in a large number of parameters to compute. To mitigate redundancy and make DCNNs more efficient, this article implements and demonstrates the concept to identifying and removing redundant or low-impact connections from fully connected layers using convolution neural network with hierarchical modular organization. The modularity of the DCNNs is built based on the cluster hierarchy of the similar image. These clusters are created based on a novel similarity metric, which measures how closely related images are to each other. The architecture uses multiple smaller DCNNs, referred to as modules, designed to progressively classify images into super clusters according to their similarity. Experimental results using popular image datasets show that the proposed DCNNs model to optimized number of operations by 49% to 99% and keeps its performance comparable.