Active Memory Learning for Mobile Sensing Systems

PROJECT TITLE :

Memory-Aware Active Learning in Mobile Sensing Systems

ABSTRACT:

A novel active learning framework for activity recognition utilizing wearable sensors is presented here. When deciding which sensor data should be annotated by the oracle, our work is distinctive in that it takes into account the constraints imposed by the oracle. Our strategy draws its motivation from the limited capacity of human beings to respond to prompts presented on their mobile device. This capacity constraint is manifested not only in the number of queries that a person can respond to within a given time frame, but also in the time lag that exists between the issuance of the query and the response provided by the oracle. We present the concept of mindful active learning and propose a computational framework that we refer to as EMMA in order to maximize the active learning performance while taking into account the informativeness of sensor data, the query budget, and human memory. We begin by formulating the optimization problem at hand, then suggest an approach to modeling memory retention, go on to discuss the complexity of the issue, and finally suggest a greedy heuristic as a solution to the optimization problem. In addition, we design a method for performing mindful active learning in batch, in which multiple sensor observations are selected simultaneously for querying the oracle. This allows us to perform active learning in a more efficient manner. We demonstrate the efficacy of our method by utilizing three activity datasets that are readily available to the public and by simulating oracles with a variety of memory capacities. We demonstrate that the accuracy of activity recognition can range anywhere from 21 to 97% depending on the size of the memory pool, the query budget, and the degree of difficulty of the Machine Learning task. When only the informativeness of the sensor data is taken into consideration for active learning, our findings indicate that EMMA achieves an accuracy level that is on average 13.5 percent higher than the case would be otherwise. This is the case. In addition, we demonstrate that the performance of our method is no more than twenty percent lower than the experimental upper-bound and is anywhere from seventy to eighty percent higher than the experimental lower-bound. In order to determine how well EMMA performs active learning in batches, we designed two different instances of EMMA, each of which could carry out active learning in batch mode. We demonstrate that these algorithms reduce the amount of time required for algorithm training, but their performance accuracy suffers as a result. One more thing that our research has shown is that incorporating clustering into the process of selecting sensor observations for batch active learning can boost activity learning performance by an average of 11.1 percent. This improvement is primarily attributable to the fact that clustering helps reduce redundancy among the selected sensor observations. We have found that mindful active learning is most useful in situations where the query budget is limited and/or the oracle's memory is not very strong. This observation highlights the advantages of utilizing mindful active learning strategies in mobile health settings that involve interaction with older adults and other populations that suffer from cognitive impairments.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here