Accelerating Neural Networks With Limited Hardware Resources by Jinglan Liu

Neural networks (NN) have been demonstrated to be useful in a broad range of applications such as image recognition, automatic translation and advertisement recommendation. But it is known to be both computationally and memory intensive, due to the ever-increasing deep structure. We consider to combine NN with selector: cluster raw input into different groups and assign every group a niche NN or a simpler model. With reconfigurability of FPGA, adopt different niche NN for different raw input. Such niche NN may fit limited hardware and realize better accuracy meanwhile.

Resource Constrained Real-time Advanced Driver Assistance System (ADAS) Application by Tianchen Wang

This research involves developing ADAS application on resource constrained hardware platform. With the help of computer vision, machine/deep learning and hardware acceleration, the goal is to get the system running in real-time while keeping the extremely high accuracy and robustness. The functions of ADAS include but not limited to lane departure warning (LDW), forward collision warning (FCW), traffic sign recognization (TSR) and pedestrian collision warning (PCW).

On-Going Project by Jie Wu

Smart grid, managed by intelligent devices, are unarguably the backbone of national infrastructure to provide a vital support for residential customers to optimally schedule and manage the appliances' energy consumption. Due to the fine-grained power consumption information collected by smart meter, the customers' privacy becomes a major security concern. With the highly accurate power consumption information, an attacker could identify customers' personal behavior patterns, such as residential occupancy and social activities. Moreover, household appliances are not homogeneous in terms of schedulability and can be classified into schedulable appliances and non-schedulable appliances. The operations of schedulable appliances can be controlled by a scheduler. On the other hand, the non-schedulable appliances must be turned on immediately upon the customers' request, the timing of which cannot be known by the scheduler ahead of time. The non-schedulable appliances, such as TV and computers, introduce operation uncertainties. Therefore, we focus an on-line appliance scheduling design to protect customers' privacy in a cost-effective way, while taking into account the influences of non-schedulable appliances' operation uncertainties. We formulate the problem by minimizing the expected sum of electricity cost and achieving acceptable privacy protection. Without knowledge of future electricity consumptions, an on-line scheduling framework (PACES) is proposed based on the only current observations by using a stochastic dynamic programming technique. The PACES is illustrated in Fig.1.

Fig.1: PACES for on-line privacy-aware cost-effective scheduling




Currently openings are available for
Ph.D. students interested in:

Preference will be given to M.S. students with strong background in maths (especially statistics or optimization) and/or C/C++ programming.

If interested, you may send an email with a copy of your CV and transcript to the email address listed above. Due to the large number of applications, not all emails will be replied.