Computer and Systems Engineering Seminar Series: Efficient and resilient learning in emerging device
Emerging device, such as ReRAM, is highly efficient to accelerate various DNN based applications, such as approximate computing. However, in device level, these emerging devices suffer from low yield due to various defects and process variations. In application level, on the other hand, the complex distributions of the input data not only incur unacceptable errors, but also degrade the invocation of accelerators. To mitigate above challenges, this talk first discusses application-level solutions to enhance the yield and resilience of the ReRAM based accelerators, and then present novel training methods to enhance the efficiency of DNN-based approximate computing.