CrowdLearner: rapidly creating mobile recognizers using crowdsourcing

Abstract

Mobile applications can offer improved user experience through the use of novel modalities and user context. However, these new input dimensions often require recognition-based techniques, with which mobile app developers or designers may not be familiar. Furthermore, the recruiting, data collection and labeling, necessary for using these techniques, are usually time-consuming and expensive. We present CrowdLearner, a framework based on crowdsourcing to automatically generate recognizers using mobile sensor input such as accelerometer or touchscreen readings. CrowdLearner allows a developer to easily create a recognition task, distribute it to the crowd, and monitor its progress as more data becomes available. We deployed CrowdLearner to a crowd of 72 mobile users over a period of 2.5 weeks. We evaluated the system by experimenting with 6 recognition tasks concerning motion gestures, touchscreen gestures, and activity recognition. The experimental results indicated that CrowdLearner enables a developer to quickly acquire a usable recognizer for their specific application by spending a moderate amount of money, often less than $10, in a short period of time, often in the order of 2 hours. Our exploration also revealed challenges and provided insights into the design of future crowdsourcing systems for machine learning tasks.