Low-Power k Nearest Neighbors Classifiers with Approximate Nanoscale Memories

The implementation of machine learning classifiers in embedded devices poses significant challenges in terms of power consumption. In some classifiers, like for example the
$k$
nearest neighbors (
$k$
NNs), the memory accounts for a significant portion of power dissipation and thus, utilizing approximate memories is an attractive option to reduce it. In this paper, the use of approximate memories to implement kNNs classifiers is considered. Efficient schemes that reduce the impact of the errors introduced by the approximate memories on the classification result are proposed and evaluated. The results show that power dissipation of memory can be reduced by 43.75% (24.77%) when using SRAMs (DRAMs), while having a negligible impact on the classifier accuracy.