1. Wireless sensing & computing towards smart homes via millimeter wave radar
Millimeter wave radar recently receives a lot of attention because of the capability of sensing and fewer privacy leaking concerns. We mainly study to apply millimeter wave radar towards indoor scenarios, for localization, counting, tracking, warning of falling down, gesture/posture/gait recognition, et al. In this topic, we not only design machine learning based models for various applications but also consider edge computing problems like performance optimization in embedded platforms and/or edge-cloud coordination. The work is cooperated with Miaomi Technologies Co. Ltd. (苗米科技)
2. Parallel Job Scheduling
A moldable parallel job could be assigned a set of computation resources (e.g., CPUs, GPUs) before execution. The more computation resources used, the less execution time is needed (e.g., t2<t1). However, the more resources used, the more data synchronization costs arise, such that the workload of a moldable job is monotonic non-decreasing when the computation resources used increase (e.g., the area t2 * r2 >= t1 * r1). With such a job model, we are studying the problem of minimization the makespan when scheduling a set of moldable parallel jobs with limited heterogeneous resources (e.g., CPUs, GPUs).
3. Accelerating DNN models towards vertical applications
Deep neural network as a typical machine learning model can solve a lot of practical problems. However, deep neural network often involves a large amount of computations that can hardly adapt many edge scenarios with limited computation resources. We are studying the problem from a system architecture view that deep neural network model is lighten and designed to be hardware-aware, and meanwhile, dedicated hardware accelerator is proposed aiming to well map the neural network.
