Memory-Aware Active Learning in Mobile Sensing Systems.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Activity recognition using wearable sensors is crucial for health monitoring.
• Traditional active learning methods often overlook the practical constraints of human annotators (oracles).
• Limited human capacity for responding to queries and memory retention impacts active learning efficiency.

Purpose of the Study:

• To propose a novel active learning framework that incorporates oracle limitations for wearable sensor-based activity recognition.
• To introduce the concept of mindful active learning and the EMMA computational framework.
• To optimize active learning performance by considering data informativeness, query budget, and human memory.

Main Methods:

• Formulated an optimization problem for mindful active learning.
• Developed an approach to model human memory retention.
• Proposed a greedy heuristic and batch active learning strategies to solve the optimization problem.
• Integrated clustering to reduce redundancy in batch selection.

Main Results:

• EMMA achieved 13.5% higher accuracy on average compared to methods using only data informativeness.
• Activity recognition accuracy ranged from 21% to 97% based on memory strength, query budget, and task difficulty.
• Batch active learning improved training time, with clustering enhancing performance by 11.1% on average.

Conclusions:

• Mindful active learning, particularly EMMA, significantly improves activity recognition by accounting for human oracle constraints.
• The framework is most beneficial with small query budgets or weak oracle memory, highlighting its utility in mobile health.
• EMMA demonstrates a practical approach to active learning in real-world scenarios with human-in-the-loop systems.