Machine Learning Based Predictive Models in Mobile Platforms Using CPU-GPU

Physiological signal based interactive systems communicate with human users in real time manner. However, the large size of data generated by sensors, complex computations necessary for processing physiological signals (e.g. machine learning algorithms) hamper the real-time performance of such systems. The main challenges to overcome these issues are limited computational capability of mobile platform and also the latency of offloading computation to servers. A solution is to use predictive models to access future data in order to improve the response time of the system. However, these predictive models have complex computation which result in high execution times on mobile phone that interferes with real time performance. With the advent of OpenCL enabled GPUs in mobile platform, there is a potential of developing general purpose applications (e.g. predictive models) which offload complex computation to GPUs. Although the use of GPUs will reduce the computation time in physiological signal based mobile systems, satisfying the time constraints of these systems can be challenging. That is due to the dynamically changing nature of physiological data which requires frequent updating of physiological models in the system. In this work, computations of a predictive model for brain signals is offloaded to mobile phone GPU. The evaluation of the performance shows that GPU can outperform CPU in mobile platform for general purpose computing.