Memory-Aware Active Learning in Mobile Sensing Systems

Abstract:

Wearable sensors enable an active learning framework for activity recognition. Our work is unique because it considers the oracle’s limitations when selecting sensor data for annotation. Humans’ limited mobile device response time inspired our approach. This capacity constraint affects both the number of queries a person can respond to and the time between query issuance and oracle response.

We introduce mindful active learning and propose a computational framework, EMMA, to maximize active learning performance by considering sensor data informativeness, query budget, and human memory. We formulate this optimization problem, model memory retention, discuss its complexity, and propose a greedy heuristic to solve it.

We also design a batch mindful active learning method that selects multiple sensor observations to query the oracle. Our method is tested using three publicly available activity datasets and oracles with different memory strengths. Depending on memory strength, query budget, and machine learning task difficulty, activity recognition accuracy ranges from 21 to 97%.

EMMA also has an average accuracy 13.5 percent higher than active learning using only sensor data informativeness. Our approach performs at most 20% below the experimental upper-bound and up to 80% above the experimental lower-bound.

We create two batch active learning EMMA instantiations to test its performance. These algorithms reduce algorithm training time but reduce performance accuracy. Our work also found that clustering reduces redundancy in sensor observations selected for batch active learning, improving activity learning performance by 11.1 percent on average.

Mindful active learning works best when the query budget is low and the oracle’s memory is weak. In mobile health settings that involve older adults and other cognitively impaired populations, mindful active learning strategies are beneficial.

Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.