Alwyn Burger, Chao Qian, Gregor Schiele, Domenik Helms
2020 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops)
IoT systems that employ AI and neural networks for processing sensor data are usually dependent on an active connection to the cloud, or restricted to simplified models and techniques. By employing a highly optimised Convolutional Neural Network that can be executed directly on our heterogeneous IoT device, we can address both these problems without sacrificing accuracy. Additionally, leveraging the cloud when unknown or uncertain samples are seen allows us to continuously improve the model. Capable of running a “normal” model trained using frameworks such as TensorFlow, we show that our system can achieve accuracy of up to 97% in medical applications such as ECGs. This is done using our device that uses less than 200mW but can locally process more than 300 heart beats per second.
March / 2020
LUTNet An energy-efficient AI network of elementary lookup tables