Towards Trustworthy Machine Learning For Human Activity Recognition
Human activity recognition is a complex field that involves understanding and classifying various human activities using sensor data. With the advancements in machine learning, deep learning, and sensor technology, there have been new possibilities for accurate human activity recognition. However, there are still challenges that need to be addressed to ensure the trustworthiness of machine learning models in this field.
One of the main challenges in human activity recognition is the effective use of temporal data. Time-series data collected from wearable sensors captures intricate aspects of human activities. To improve the classification performance, a new temporal ensembling framework has been proposed. This framework utilizes the temporality of the data and trains an ensemble of deep learning models. By accommodating various window sizes for time-series data, it enhances classification accuracy while preserving temporal information.
Another important aspect of human activity recognition is the reliability of predictions. While accurate prediction is crucial, the reliability of those predictions often goes unnoticed. To address this issue, the aforementioned temporal ensembling framework is used for calibration and uncertainty estimation. It provides well-calibrated predictions and detects out-of-distribution activities, thus improving the reliability of human activity recognition models.
In addition to reliability concerns, privacy is also a major issue in machine learning models used for human activity recognition. The utilization of sensitive activity data during model training raises privacy concerns. While techniques like federated learning can provide some degree of privacy protection, they lack quantifiable privacy metrics tailored to individual preferences. To tackle this challenge, a differentially private framework for time-series human activity recognition has been introduced. It quantifies privacy and enhances privacy preservation. Furthermore, a collaborative federated learning framework allows users to define their privacy preferences, advancing privacy protection in human activity recognition.
These novel solutions address the challenges of utilizing temporal data effectively, improving reliability, and preserving privacy in human activity recognition. They enhance the overall performance and trustworthiness of machine learning models employed in this field. With these advancements, machine learning in human activity recognition can be used in realistic and practical scenarios, bringing us closer to trustworthy machine learning in this domain.
In conclusion, the field of human activity recognition has seen significant progress with the integration of machine learning, deep learning, and sensor technology. However, challenges such as effective utilization of temporal data, reliability of predictions, and privacy preservation remain. The proposed solutions aim to address these challenges, promoting improved classification, reliability, and privacy in human activity recognition.
