DeepSense

Recently, a branch of machine learning algorithms called deep learning gained huge attention to boost up accuracy of a variety of sensing applications. However, execution of deep learning algorithm such as convolutional neural network on mobile processor is non-trivial due to intensive computational requirements. In this paper, we present our early design of DeepSense - a mobile GPU-based deep convolutional neural network (CNN) framework. For its design, we first explored the differences between server-class and mobile-class GPUs, and studied effectiveness of various optimization strategies such as branch divergence elimination and memory vectorization. Our results show that DeepSense is able to execute a variety of CNN models for image recognition, object detection and face recognition in soft real time with no or marginal accuracy tradeoffs. Experiments also show that our framework is scalable across multiple devices with different GPU architectures (e.g. Adreno and Mali).