My primary research area is Efficiency-Quality Tradeoff through Imprecise Computation. Many real-world applications are constrained by power, area or performance (efficiency) but can tolerate certain degree of quality degredation. For example, it is sometimes difficult to implement a complex image processing algorithm at a rate to process a video stream in real time. However, if we allow the algorithm to occassional produce inaccurate results or to perform approximate computation, the performance requirement can likely be met.

Sacrificing correctness might not sound appealing to many people, however if we think carefully about the intended use of many applications, do they really need to be THAT accurate? Take human-interacting applications for example. If an error is not discernible by human percecption, why should we design the hw to always operate correctly? If an approximate design produces error unrecognizeable by humans but at a much lower implementation cost, isn't it a better choice?

My focus is to develop a methodology to optimize the efficiency metrics of a HW design given a set of quality constraints. Although efficiency-quality tradeoff can be achieved in both HW and SW level, I believe HW provides more flexibility in terms of the number of tuneable knobs, less overhead and usually yields a better design than SW.