As the use of radio frequency identification (RFID) technology becomes widespread, one of the actively pursued research areas has been integration of smart sensors with RFID tags. These miniaturized devices when queried, transmit an identifier, subject to a specific condition of their environment. Applications of the sensors range from surveillance to biomedical systems. In this paper we will present a floating gate classifier that detects patterns of interest and enables or disables RF transmission on the tag. These classifiers consume minimal amount of energy and can be directly operated by scavenging power through inductive coupling. Applying online learning and calibration techniques to the classifier can compensate analog imperfections and mismatches in sensors. Results obtained from a prototype fabricated using standard 0.5μm CMOS process, are presented and the utility of learning on silicon is validated by demonstrated improvements in detection rates.