Efficient Computer Vision Inference using Modular Neural Network Techniques.

Deep learning, especially neural networks, has become significant achievements in the realm of artificial intelligence (AI), including areas such as natural language processing, computer vision, and speech recognition. Furthermore, the impressive performance of deep learning models is often followed by a significant drawback: their high computational complexity. This limitation poses a challenge when deploying these models on resource-constrained devices like Internet of Things (IoT) devices. In this paper, we implement and demonstrate the concept of modularity into Deep Neural Networks (DNNs) to reduce the redundant operations and minimize the loss of its performance.