Fault tolerance of self-organizing maps

Abstract

Bio-inspired computing principles are considered as a source of promising paradigms for fault-tolerant computation. Among bio-inspired approaches, neural networks are potentially capable of absorbing some degrees of vulnerability based on their natural properties. This calls for attention, since beyond energy, the growing number of defects in physical substrates is now a major constraint that affects the design of computing devices. However, studies have shown that most neural networks cannot be considered intrinsically fault tolerant without a proper design. In this paper, the fault tolerance of self-organizing maps (SOMs) is investigated, considering implementations targeted onto field programmable gate arrays, where the bit-flip fault model is employed to inject faults in registers. Quantization and distortion measures are used to evaluate performance on synthetic datasets under different fault ratios. Three passive techniques intended to enhance fault tolerance of SOMs during training/learning are also considered in the evaluation. We also evaluate the influence of technological choices on fault tolerance: sequential or parallel implementation, weight storage policies. Experimental results are analyzed through the evolution of neural prototypes during learning and fault injection. We show that SOMs benefit from an already desirable property: graceful degradation. Moreover, depending on some technological choices, SOMs may become very fault tolerant, and their fault tolerance even improves when weights are stored in an individualized way in the implementation.