The use of neural networks in mobile sensing is becoming increasingly popular. Neural networks, however, bring high performance at the cost of resource intensiveness and high power consumption. To mitigate that, we can use a variety of optimization techniques, which reduce computational demands, but also reduce accuracy. Since mobile sensing is inherently dynamic, using only one static level of optimization is not ideal. In this work, we propose an approach to dynamically selecting the level of optimization during real-time classification. By doing so, we can use a simpler model for classifying examples that are easy to classify, thus saving some energy and a more complex classifier for more difficult examples, preventing accuracy drop. We implemented the approach on an activity detection application and showed that by using dynamic optimization selection we can achieve higher accuracy than would be possible by statically selecting the same optimization level. We also implemented the activity detection pipeline on an actual mobile device and confirmed our results with an experiment of our own.
|
