Extending naive Bayes with precision-tunable feature variables for resource-efficient sensor fusion

Resource-constrained ubiquitous sensing devices suffer from the fundamental conflict between their limited hardware resources and the desire to continuously process all incoming sensory data. The data's representation quality has an immediate impact on both aspects. This paper strives to enable resource-aware and resource-tunable inference systems, which are capable of operating in various trade-off points between inference accuracy and resource usage. We present an extension to naive Bayes that is capable of dynamically tuning feature precision in function of incoming data quality, difficulty of the task and resource availability. We also develop the heuristics that optimize this tunability. We demonstrate how this enables much finer granularity in the resource versus inference accuracy trade-off space, resulting in significant resource efficiency improvements in embedded sensor fusion tasks.