Recomputing by operand exchanging: a time-redundancy approach for fault-tolerant neural networks

Abstract

The use of neural networks in mission-critical applications requires concurrent error detection and correction at architectural level to provide high consistency and reliability of system's outputs. Time redundancy allows for fault tolerance in digital realizations with low circuit complexity increase. In this paper, we propose the use of REcomputation with eXchanged Operands-an approach based on operands' rotation-to introduce concurrent error detection and correction, when timing constraints are not particularly strict. Different architectural approaches for neural design are considered to match the implementation constraints and to show the versatility of the proposed solutions.