Bipolar Vector Classifier for Fault-tolerant Deep Neural Networks
TimeWednesday, July 13th2:37pm - 3pm PDT
Location3000, Level 3
ML Algorithms and Applications
DescriptionFor reliable operation of DNNs, many researches are proposed to improve the fault tolerance of DNNs. Although they improve the average of DNNs' fault tolerance, they didn't prevent the worst case of fault situation, especially in classifier of DNNs. In the worst case, only one bit fault in the conventional classifier can cause the significant accuracy drop. This paper proposes a novel fault tolerant classifier method which provide equivalent accuracy and can be incorporated easily with other techniques. The proposed method guarantees the accuracy in any worst fault situation up to 10^(−3) BER in classifier.