Optimal weight storage improves fault tolerance of SOMs

Bio-inspired computing principles are considered as the 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 properties of Self Organizing Maps (SOMs) are investigated, considering a general fully parallel digital implementation on FPGAs. To assess their intrinsic fault tolerance, we use the bit-flip fault model to inject faults in registers holding SOM weights, and the quantization and distortion measures are used to evaluate performance on synthetic datasets under different fault ratios. Experimental results are analyzed through the evolution of neural prototypes during learning and fault injection. We show that SOMs become very fault tolerant when their weights are stored in an optimal way.