Confidence-Calibrated Human Activity Recognition.

Area of Science:

• Human activity recognition (HAR)
• Machine learning
• Wearable sensor technology

Background:

• Wearable sensors and deep learning (DL) excel in activity recognition (AR) for health, sports, and care.
• Current DL models prioritize accuracy over reliable confidence calibration, leading to overconfident predictions.
• Uncalibrated probabilistic estimates hinder the trustworthiness of DL models in critical applications.

Purpose of the Study:

• To introduce a novel ensembling method, deep time ensembles, for calibrated confidence estimates in neural network architectures.
• To address the unreliability of overconfident predictions from current deep learning models in activity recognition.
• To enhance the trustworthiness and practical applicability of wearable sensor-based activity recognition systems.

Main Methods:

• Proposed deep time ensembles, an ensembling technique for neural networks.
• Trained an ensemble of models using temporal sequences from varying window sizes on time series data.
• Averaged predictive outputs from the ensemble to generate calibrated confidence estimates.

Main Results:

• Demonstrated significant improvement in confidence calibration across four benchmark Human Activity Recognition (HAR) datasets.
• Reduced expected calibration error (ECE) by at least 40%, indicating more reliable likelihood estimates.
• Outperformed state-of-the-art classification on WISDM, UCI HAR, and PAMAP2 datasets, and matched state-of-the-art on the Skoda dataset.

Conclusions:

• Deep time ensembles provide calibrated confidence estimates, enhancing the reliability of wearable sensor-based activity recognition.
• The proposed method improves both the calibration and classification performance of deep learning models for AR.
• This advancement offers more trustworthy probabilistic outputs for diverse applications leveraging human activity recognition.