PARTITIONING AND MANAGING ARTIFICIAL INTELLIGENCE IN COMPUTING CONTINUA FOR SMART EYEWEAR
Artificial Intelligence (AI) applications (frequently based on Deep Neural Networks - DNNs) are increasingly popular as they seamlessly integrate into end-user devices, enhancing overall quality of life. These applications span diverse domains such as healthcare, entertainment, industry, and engineering.In recent years, there has been a growing focus on the development of Smart Eye-Wear (SEW) to enhance user experiences based on specific usage domains. However, SEWs face constraints related to computational capacity and battery life. This research investigates SEW technology, proposing algorithms to minimize energy consumption and reduce 5G connection costs while maintaining a high Quality-of-Service (QoS).To achieve these goals, a management framework is under development to offload some of the DNN computations to the user's smartphone and/or the cloud, thereby leveraging the potential for DNN partitioning.Different optimization problems are formally modeled as a Markov Decision Processes (MDPs) considering different AI application scenarios, and various Reinforcement Learning (RL) algorithms are implemented to provide solutions to those MDPs. Performance evaluation will take into account factors such as the variability in 5G and Wi-Fi bandwidth, as well as latency within the cloud. A real prototype is also considered to assess the training results.Preliminary results are very promising demonstrating that RL agents are able to adapt the application at runtime optimizing SEW energy consumption, reducing 5G connection costs, and ensuring QoS.
Back to Alumni