Quantization Error and Accuracy-Performance Tradeoffs for Embedded Data Mining Workloads

摘要

Data mining is the process of automatically finding implicit, previously unknown and potentially useful information from large volumes of data. Embedded systems are increasingly used for sophisticated data mining algorithms to make intelligent decisions while storing and analyzing data. Since data mining applications are designed and implemented considering the resources available on a conventional computing platform, their performance degrades when executed on an embedded system. In this paper, we analyze the bottlenecks faced in implementing these algorithms in an embedded environment and explore their portability to the embedded systems domain. Particularly, we analyze the floating point computation in these applications and convert them into fixed point operations. Our results reveal that the execution time of five representative applications can be reduced by as much as 11.5× and 5.2× on average, without a significant impact on accuracy.